> Kimi Work is a Local Agent designed for deep workflows. It mounts your local folders, navigates the web autonomously via WebBridge, runs Python code in the background, and executes scheduled tasks.
It's clearly a dupe of Claude/Codex products (Codex especially, styling-wise), but I think Kimi's goal here is simply to appear on feature-parity with bigger labs. I doubt they'd want to invest much into designing a UI for "the future of agentic work" from the ground up.
Their bread and butter and claim-to-fame will continue to be low-cost, near-frontier models, for which they need to at least appear to have some app layer to plug into.
And, in some ways, duping is a great way to show people that even the app layer can be commoditized.
Some paradigms are simply ergonomic. Sometimes there's only one UX people will accept.
No one would buy a car steered with pedals; even if that is better (hey planes do it!) - every car maker is "copying" steering wheels and putting them on their cars. Is it controversial? If not why not?
I highly recommend to anyone that uses them a lot - try to clone your own. I’ve got mine fully formed into how I want it, with features I haven’t seen anywhere else yet and with integration into everything I work with. We’ve lived in a world where every app had to fit enough people well enough - but now we live in a world where an app can be molded to be absolutely perfect for just you!
Looks like the thing I am building but of course mine will be slightly different. It would be great if we could all meet to show off and discuss our AI stuff.
The problem of the sota models clamping down on third party harnesses should be stated upfront. Getting a SOTA model in a custom harness requires API pricing or risking an account ban, correct?
This preliminary change in cost may not make it worth it for people depending on which is their "favorite agent". Especially as the default harnesses for said agents continue to improve...
Presumably people want to hear the opinions of other people and still believe that they might learn more from the their experience than by exclusively interacting with AI models?
The alternative of that first impulse should be what exactly? Telling AI to research Medium AI drivel articles (since nobody seems to be posting long/medium form content anywhere else these days) to figure out what’s the best approach to building an AI harness?
BTW just my personal observation but I generally find AI notoriously bad at building any AI applications compared to standard software (which is not surprising given the scarcity of high quality training data)
It's pretty simple nowadays if you know conceptually how they work. Running the LLM calls in a loop with tools is an agent. You only need 10 or so basic tools to accomplish nearly anything, and you can build a dynamic skill system from that. Look at https://github.com/patw/pengy, ignore the app look at the spec.md file, feed that to your current agent of choice and make your own version. Use whatever tech stack or UI you're comfortable with. Change some of the choices in how it works, so it fits what you want to work.
It would have been much easier to clone six months ago but today Codex/Claude desktop apps are so feature packed it seems it wouldn’t be worth it.
For example, the mobile/remote control feature I find very useful but not sure it would be worth cloning. If you aren’t replicating every feature you aren’t exactly “cloning” it, you’ll just end up with a crappier version with a fraction of the features of the real thing.
Actually unless they’ve changed substantially Kimi kind of does this? It’s not as advertised/focused on and the API is a little nonstandard making it hard to plug some things into, but, unlike Claude they’ll actually give you an API key you can use directly as part of your unmetered subscription. I played with it for a month or two out of curiosity, it ls not a bad value though a bit brittle and stuff
The Chinese api prices are already so cheap that that is arguably not needed.
You can also get $10/month opencode go (which officially can be used with any harness) with basically every major open weight model running inference in a place that I deem more trustworthy than the subject-to-Chinese-law inference options (but research this for yourself, there are some asterisks). This is what I use (in opencode) when I want an agent for personal use. Even if I'm being fairly intense with my usage by my standards anyway), I haven't managed to run out (yet), although I usually use the cheaper models (deepseek v4 is my favorite, and I've used both flash and pro).
From the FAQ:
> The plan is designed primarily for international users, with models hosted in the US, EU, and Singapore for stable global access. Our providers follow a zero-retention policy and do not use your data for model training.
And:
> Yes, you can use Go with any agent. Follow the setup instructions in your preferred coding agent.
Isn’t OpenCode just reselling Deepseek’s tokens over the same API? For v4-pro the prices from other providers on OpenRouter are way higher so what else could it be?
Opencode Go already seems extremely unprofitable if people use up at least 30% of their monthly tokens so its hard to imagine they’d go with several times more expensive inference provider yet show Deepseek pricing on their page.
American labs will do fine with monthly subscriptions costing $20-300/month. They will not do fine for heavy basic LLM API calls -- it is for this that the cheap API produces substantial savings. Specialized niche provisions, e.g. excellent tool use, reliable structured output, reliable long context, cached and uncached web search, document reads, computer use, integrations, etc. also will take time for the foreign services to be good at.
So much is invested in the huge American AI companies that China undercutting them with clones that work almost as well can actually hurt the US economy.
They're at the point where copying can get people 90% of what they want anyway, the moat is drying up.
If you remove the word "AI" from your post it also describes the last 20 or so years of US/Chinese economic competition. They are real, serious competitor but also there is a lot of value in that last 10%. Some buyers need the best and you can charge a premium for it.
The speed in which a copier can catch up to a leader is faster than every before, but everything is faster these days. I think the threat is if can these labs catch up and pass American labs while keeping costs low, but I don't see that happening.
>Some buyers need the best and you can charge a premium for it.
Yes, and that is how you get a top-heavy k-shaped economy I believe. It is harder and harder to sell premium A+ products, because most people can only afford what works. If my old ice cream vendor charges me 25% more but puts chocolate sprinkles on it i'm going for the non-sprinkle version.
> the last 20 or so years of US/Chinese economic competition
Only with some industries, mostly electronics and light manufacturing. The rest of the history of software products doesn't look like this at all. China cloned Amazon and Meta fairly well (but only within their local/captured markets) but not Microsoft or Google and absolutely not Apple. Entertainment and culture exports are still effectively zero (like, what, three blockbuster films?) where the US, Japan and Korea define pop across the globe. Heavy industry is mixed, with decent success in consumer vehicles now but little purchase in stuff like heavy equipment or shipbuilding.
Basically it's complicated, and not really a point well suited to that kind of generalism.
> China cloned Amazon and Meta fairly well (but only within their local/captured markets)
Americans order Chinese products from SHEIN and Aliexpress type websites.
So China did penetrate the US ecommerce market, but they are not taking on Amazon directly.
People are buying from the Chinese and reselling on Amazon, which has a good logistics network that Chinese companies can't possibly build in the USA.
> absolutely not Apple
You need to look up Oneplus to see how many people were using their devices till they announced exit from Western markets.
The Chinese Android phones are essentially replacement for a big chunk of Apple ecosystem.
If these cheap mobile devices didn't exist - Apple could possibly be selling many more devices.
> China cloned .... Meta fairly well
TikTok is giving Meta a run for their money in USA.
They had to be restricted via threat of regulation to offload stake in TikTok USA.
> Microsoft
They didn't compete with Microsoft directly as entering Western Enterprise market is hard for the Chinese.
Even Google can't beat Microsoft there till now.
Not British, but taking UK out of the definition of pop, is a bit too rough a cut. Also France. Look at the queues outside pop-glamorous stores. And was pop-art a pop artefact? Is TikTok redefining pop? Movies is an old measure, and even US has fallen with just franchise dominating. I think as a globe we are much more integrated and co-dependent, occasionally also in bad ways, than we appreciate.
I'm talking about economic activity more than the passports of the individual artists. And no, TikTok is a software product (a social media platform), not popular culture. There's absolutely popular culture being created and driven there but it's 100% NOT coming out of China. Are you high, as it were?
TikTok competes with Meta, not Olivia Rodrigo or BTS.
I am not high, I have kids, and can tell you they barely care about Olivia Rodrigo or BTS. You dispel passports as irrelevant for one leg but base your argument on them for the other. The contradiction is sufficient.
> Entertainment and culture exports are still effectively zero (like, what, three blockbuster films?) where the US, Japan and Korea define pop across the globe.
And this is really sad, because they have thousands of years of cultural heritage and also plenty of great films. Somehow, they've failed at exporting this completely.
> Heavy industry is mixed, with decent success in consumer vehicles now but little purchase in stuff like heavy equipment or shipbuilding
Uhh, China does over half of the world’s shipbuilding.
China is also responsible for 25% [1] of the world’s heavy equipment manufacturing and is considered to have double[2] the CAGR of the global average for heavy equipment manufacturing.
the difference is that the US economy is currently over-investing in AI to the point where it's a significant portion of the economy, it's the biggest competitive opportunity in quite a while
China isn't undercutting anything. Kimi has to be hosted somewhere and is extraordinarily expensive to operate at Fable-levels of capabilities. The moat is (and will always be) the cost to actually operate these models, and that cost will continue to soar, it will not reverse, and the hardware will never keep up such that at-home use somehow becomes economical.
It's like releasing a cutting edge search engine, proclaimed to be on par with Google. Oh yeah, I forgot to mention, you need to actually run it somewhere. One of the core moats for Google has been the cost to scale and compete globally with a search engine, which only a few entities on earth could even attempt. Somehow the HN crowd forgets this every single time. The same claims were endlessly made regarding Uber, as though the app was the hardest part of competing with Uber. Here's a hint: the model isn't the only super hard part of the total process at this point. Try running something akin to GPT 5.6 for a billion users, see how much that costs.
The winners will be the owners of the subscriptions and API usage. That's going to be OpenAI and Anthropic as of now, possibly Google will stay in the fight with their scale advantage focused on the consumer. It will never be the open weight models out of China. Who is going to invest the trillion dollars into infrastructure needed to run the latest version of Kimi 3-5 years out, for a billion users? There's only one entity that would ever consider it, and that's China domestic.
And the real money (for a sustainable business model allowing for constant massive reinvestment & infrastructure maintenance) is in pairing up advertising, which is another spot where the open weight models from China will never compete. It doesn't matter if the HN crowd hates this outcome, nothing will stop it from being the future core of mass consumer usage. And the business use will stay with subscription + API use as well: the best option for that today is OpenAI and Anthropic.
>Who is going to invest the trillion dollars into infrastructure needed to run the latest version of Kimi 3-5 years out
Everybody that is concerned about their data ownership and residency, and how the US political system is behaving right now. There is a huge incentive to cut off ties with API and subscription providers, especially those based in the US.
People endlessly claimed the same thing about DeepSeek upon its release. they claim it every single time a very large cutting edge model is released out of China. It's a combination of propaganda, some truth, and the faux open culture desperation to see big tech get destroyed (somehow better if big China takes its place instead).
(that's just one, there are others... and several are claiming that those saying "yes sure it's just as good, but it uses more tokens" ain't really true either).
And I don't think anyone is excited about China releasing anything: I think people like the idea of open-weights.
All bikes look the same, all cars look the same, all cameras look the same, all bridges, all dams, hell all smartphones look exactly the same! The whole goddamn point of UX is that you make something that is usable, and hence familiar to your users.
People who say "I was making GPUs before them, and they built their own GPUs that behave like GPUs so they copied me" don't understand how stuff is built. The hard part in engineering is not "imagining the high-level concept", it's actually making it work. Even if you can open electronics and study it to design and build your own, it requires a lot of expertise. And if you manage to do it, then you are an expert, period.
I'm sure many people find comfort in saying stuff like "You know what? I was a bit of a visionary because when I saw the first personal computer, I imagined that it would be cool if it could be orders of magnitudes faster and smaller such that it fits in a pocket". As if that was the hard part.
Typically, too many people spend a lot of time talking about UI and stuff like "unpacking" as if that was the most important part. But usually that's because they have no idea about how it works under the hood, so the only thing to discuss is whether the round edges are great or not.
Yeah, but maybe Kimi doesn't pretend to know better than I do. Like, I just had a task today where I had a silly form where I wanted to copy in a signature from one document to another document. And, you know, I could have just used some other dumb tools, but I decided, hey, let me us ask Codex to do it. But it started lecturing me on how we can't do this kind of stuff. And it was just pure silliness. So hopefully Kimi Work doesn't try to pretend that it knows better and lets me decide on ethics.
This blog post here summed it up the best. These are tools, so let them be tools.
I'd prefer it uncensored too but this objection was insightful in 2023. Everyone and their dog knows they're censored on topics sensitive to their country, much like western models are censored for western woke sensitivities. Nobody cares anymore as long as the reasoning is good and the price is low
kimi has been particularly shameless in copying codex 1:1, wow.
Like, I get it, before condex there was conductor and a ton of other guis but this looks like they just copied the code.
Also, their privacy disclosure is incredibly misleading:
> How does Kimi Work protect my privacy when accessing local files?
You have absolute control over your files. The built-in Ask before acting safeguard means Kimi will prompt you for explicit authorization before it modifies, overwrites, or runs code within your local directories. Nothing happens without your consent.
They fail to mention that is has unfettered read access to your files. ie, without your consent.
You have almost no control with Codex or Claude in that you have to trust them not to break their own rules.
You have almost no control with Kimi either, they can also say one thing and do another. You will not be able to sue a Chinese company if it breaks the law.
I can think of no faster way to hand over your trade secrets or intellectual property than to run something like this.
People will use Kimi anyway though because they are naive and don't value those things.
“How does Kimi Work protect my privacy when accessing local files?”
It doesn’t protect your privacy. It’s like asking, “how do I know you’re not spying on me?” And getting the reply “we cannot physically enter your house.”
>It doesn’t protect your privacy. It’s like asking, “how do I know you’re not spying on me?” And getting the reply “we cannot physically enter your house.”
In light of the recent grok build snafu, "Nothing happens without your consent" seems like an upgrade. Moreover what are they supposed to do here? By that logic should any sort of non e2e email/storage provider not be able to put "How does [product] protect my privacy" in their faq?
Looking through their privacy policy, it looks like the bigger issue is that they are (or refuse to refute) training on your data/prompts. That's probably what people should be complaining about rather than complaining about how it "has unfettered access to your files", when that describes all coding harnesses.
"How does Kimi Work protect my privacy" can be interpreted in two ways. The first is "How does Kimi Work protect my privacy from Moonshot?" They did not answer this question. The second is "How does Kimi Work protect my privacy from other third parties?" They are answering this question by saying, we won't upload your files to some third party server without your consent.
It is, but I really think this project would be better as open source. They are already all-in on open for the most valuable part (the frontier model)! They might as well make the harness fully transparent so that questions about what's being accessed and transmitted can be avoided.
I personally wouldn't use any agentic harness from any company, American, Chinese, or otherwise, that is closed source. The Grok fiasco shows why.
Sorry, I meant that’s not desirable behavior that the user might expect. I understand that coding agent tools tend to play fast and loose with access because it’s not their own privacy/security at stake.
Not sure what that's about, I was just copying the wording from OP. Their harness probably has the same restrictions as most other coding harnesses. That is, the read tool gives all project files by default without prompting, and everything else requires manual approval (or opting into auto-run all). If you decide to run it from your home directory however, then your "project" is all your files.
Sharing among AI friends what’s the problem, scraping the Internet or hiring new employees between companies who like to share? It’s a tradition among the AI model makers…
It gets tricky when you want to say “allow this particular tool to have a little more access.” Like allow the codex client to access the GPT API but not some random other site the LLM learned about. Or like letting git write to .git to commit changes but don’t let something install git hooks. Or allow downloading certain packages/versions from a site like PyPI, or allow certain operations on certain GitHub repos in a way that doesn’t line up with how you can configure an access token.
You mean other than the zillion dollars of infrastructure you need to run these systems for a billion users. And the extreme cost to sustain that operation day to day. And the cost to upgrade the infrastructure forever.
Remind me which entities (govt or corporate) are going to subsidize that forever. Remind me which entities are going to front the cost just to scale up.
By that premise - wherein you forget everything other than the first ingredient (the model) - Google has no moat either (got $100 billion dollars lying around?). Nor does Uber (have tens of billions in your bank account to clone their business?).
I'm certain building an app (Uber), or building a search engine (Google), is the super duper hard part.
Slapping a GUI on Linux will kill Windows, any day now.
Hyperscalers would generally be expected to sell compute to anyone who is willing to pay assuming that supply/demand for datacenter hardware evens out at some point that should become true again.
I wish they had some kind of US hosted option, NDR policy, or something like that. I love the price and power, but I have concerns around data sovereignty and leaking IP to an overseas company.
As soon as the K3 weights are published on July 27th, there will be many US providers hosting the model. I realize model hosting doesn't really apply to kimi work specifically, but in terms of accessing K3, there will be US options likely in 7 days. Just look at OpenRouter to see the providers hosting K2.6 https://openrouter.ai/moonshotai/kimi-k2.6.
Looking at these pages reminds me of browsing jellybean components on DigiKey. Amazing for the consumer, but one does wonder: What are all these DotAI infra companies’ plans to justify any of their valuations if they’re just selling commodity compute on OpenRouter?
Same. No way I am uploading data to a PRC company. It's not just an IP issue ... there's a very high likelihood such companies will be restricted in the United States, even if they try "Chinawashing" stunts like Bytedance.
Though it's ever so slightly broken (the sub-agent tracking stop button does nothing, some toggles in the Settings are not toggleable and so on, really slows down with long conversations to the point where keyboard input lags by seconds, think 600+ turns).
Personally, I still liked ZCode better and will probably write a blog post about my experience with K3 too, the harness might need some more work, but the model is really, really good.
Say's who? Why? I honestly don't get the problem. OpenAI just did what you said with the Mac app and were forced to roll it back because of the uproar. A chat box is simple, relatable, easy to use. K.I.S.S. is a thing
Says everyone who's not using this stuff. It's an empirical argument. Is anyone actually using this for work? How long has it been since Cowork was released?
It's such an inadequate tool for work it reeks of SF hubris to think you can just slap a chat input box onto a PC and replace 90% of white collar jobs.
it's pretty great for coding because the CLI provides the model with all the context it needs and with a way to verify the result of its actions via linting, compiler messages and program output
Models need richer context to do well. They also generally need a verifier, but better context is already an improvement. So to fix it, you need to rethink UIs to be agent-first.
I'm not saying Kimi Work and its peers are entirely useless. I'm just saying they're a hack that puts a pretty low upper limit on agent capabilities and speed.
It's like this in the early days of any technology. It reminds me of the early days of the internet before everything became hyper optimised algorithms.
It's like the late stages of a monopoly game - it's not fun for anyone except the main winner.
> Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritizing compute for current members. Existing subscribed users are not affected.
> We're adding capacity as fast as we can and will reopen new subscription spots in batches.
That might even be a beneficial move, creating (well, actual, not artificial) scarcity works well to draw interest usually. Curious how this one plays out.
Looks like this can use webBridge, run Python and shell code in the background. Cowork can take over application UI as well which is still a differentiator (and a higher security risk) unless kimi offers it elsewhere.
Is there a link anywhere for a docs site for this? Really curious to know what "deep data sources" they have set up for the Native Global Market Data feature.
while it's true the model layer is getting commoditized, there is absolutely no way we are migrating any serious enterprise workflows to Kimi work. i think chinese models will continue to have an adoption problem.
If history is anything to go by, in the next week we'll see a distill of GPT 5.6 or Mythos that is not as powerful as either, but pokes its head just in front of Kimi K3.
Agents are quite capable for any work involving CLI (not TUI) applications.
Seems to work for me to manage my emails. I have Proton Mail bridge and `mbsync` dumping all mail to a local folder, with an inotify watch script to react to new emails and classify them as either needing my attention (ignore) or something that does need my attention (send me a notification).
I've been looking for a good CLI spreadsheet editor
I'm kinda confused given that they're from China. The Chinese Internet is very different from the western Internet. The web is way smaller, and a lot of useful content is behind app-based walled ecosystems such as Wechat, Douyin and Xiaohongshu. This makes web automation much less useful. And I don't see them marketing something like Wechat automation?
Has anyone cancelled their $200 ChatGPT subscription and started using Kimi? It'd be an interesting metric to measure. I for one have no motivation to use this model. Nonetheless, I'm always happy to see more competition.
I've been subbing to Claude for 2 years and somehow the questions of whatever happened in Tiananmen (I know already anyway) have never once came up, but Claude (especially Fable) is now famous for refusing to work on things people want. And even pushed back against me when I was doing asking to research some people famous (infamous) in Israel-Palestine conflict. I really don't get why these "chinese gotchas" are keep coming up in every thread, they are mostly entirely inconsequential to getting actual work stuff done while same can't be said about Fable or others. Once, Gemini refused to help me make an infographic about women in STEM because the specific woman I was doing research about has died a few years ago, other models had no such qualms. All these "gotchas" are quite arbitrary.
I can second this. I was using Fable for some research and fable said it wasn't going to do it and downgraded to opus. Switched to K3 and continued my work. And it worked really well for what I was doing. You can easily switch claude code to use K3, it took like 30 seconds.
So for your Tiananmen questions go to Fable, and for the Fable-nerfed subject matter go to Kimi (or GLM, as Hugging Face were forced to do when they were recently hacked and the nerfed US models refused to analyze it).
It rubs me the wrong way that they have to copy everything, the UI is almost exact copy of Codex. The entire mind set of just copy the best is holding them back.
It's not about copy the best, it's about reducing friction for someone coming from a competing product.
Hence why everyone uses the same play, stop etc "buttons" / signs on "mediaplayers" whether virtual or real. [1]
Unless your user really trusts you / doesn't have an alternative reducing friction is better than being subjectively better [2]
[1] triangle, square etc...
[2] bottom placement versus top placement of browserbar on mobile phones... From an ergonomics perspective way better but from a UX experience worse... Hence chrome changed top to bottom but was forced to change it back on android, while apple forced their users from top to bottom and didn't have to move it back
Give them time, they'll work out their (Chinese UI/UX) groove just like they're working out EVs in the Chinese market. Copy first, improve and adjust later.
I have a UI I developed from first principles about 10 years ago, I use it to replace my os, its like my personal do everything emacs os but not at all emacs, and its funny to see other Uis iteratively convergence on it as context finally gets its due.
Nobody actually producing hard goods, real world value and actual software are using these cute tools. Its completely worthless to people actually doing things.
Will be cool until the AI goes ape shit and starts deleting all your files, wipes out your corporate file share, but not before uploading it all online.
Not saying Kimi will do this, but it's perfectly reasonable to expect this to happen, because there is really nothing anyone can do to prevent it. No, this is not meant to go in an isolated VM on a Linux box in a cheap VPS somehwere. This is meant for regular home users and corporate users to install on their primary system where all their personal and corporate data is available.
The above is a PSA (public service announcement) not just for Kimi, but also for GPT and Claude. If you don't understand how it's impossible to prevent the above from happening, then you should never use one of these tools.
For the past 3 years, many software engineers were coping that AI (particularly LLMs) was only going after juniors. Given the documented history of mass layoffs related to AI from 2024 to today, it is clear that everyone including seniors and staff level folks are just as affected.
So let's just say, if your whole identity is for being a typical software engineer no matter the seniority and nothing else, perhaps it is time for you to reconsider options as of course 100% avoiding AI is futile at this point.
There will still be new jobs, but there will be no need for more excess people, even under Jevon's paradox. Software engineers were just the test bed for this and now there is an oversupply of SWEs without jobs in the market and will have to use AI tools like Kimi Work and Claude Cowork to be employable.
I have never been more sure about the viable future of software engineering as a career than I am after experiencing everything all the way up to and including Fable.
First of all to reply to your deleted comment, yes this is completely relevant to the article.
As token costs drop significantly (thanks to local models and hosted Chinese models.) employers will consider using cheaper models like Kimi Work and others, even if they are already using expensive ones like Claude Cowork. They won't be able to operate without it.
Secondly, do I actually believe this? Of course, I already said that there will be more mass layoffs because of AI years ago. [0]
Who cares. Competition is good for the user even if it's built on plagiarism. The other site was probably designed using an LLM anyway. LLMs can in some aspects converge wrt their outputs.
It's a mistake to make this such an obvious copy. It will feed the narrative that Chinese entities cannot be trusted by the West. They could have varied it even slightly. It's one thing with hardware because customers don't need to trust suppliers any further than they can throw the box. But these labs seem to be battling for enterprise knowledge work-share. Trust is paramount, and they're already behind on that front. Own goal.
Funnily enough, you will have to pretty accurately copy the laws of our universe to make an Apple pie. Even small changes will change the properties of atoms, and hence all of chemistry and biology. And there won't be any apples or wheat.
Eyebrows will raise at Anthropic / OpenAI, perhaps. Why are they charging 1000x more for the exact same business?
The 'distillation is theft!' shtick won't last long. In fact, unlike Anthropic/OpenAI, these Chinese labs actually had to pay for that 'stolen' training data.
> Why are they charging 1000x more for the exact same business?
The implication they want to avoid, I'm suggesting, is because the CCP isn't plugged into and funding those entities' activities at a loss. Or they're lying about ZDR, etc. etc.
GLM 5.2 and Kimi 3 both had huge API pricing jumps (GLM most expensive chinese model, by a lot, then then Kimi 3 a lot higher than that). the cost advantage is rapidly decaying. the oft-cited cost per task makes Kimi look good vs Anthropic but even, or slightly vs OpenAI
Once the OpenAI/Anthropic-driven hype dies down people with decision-making power will be taking a look at the numbers and look at real productivity metrics. And they'll see it makes little sense to pay so much when competition is around and similarly good.
For fear that the data they're entrusted with will be copied or that there is some nefarious reason like state sponsorship that makes the activity economical?
Meh, we should just ignore that kind of propaganda. We copy things too, we use a system to protect IP because copying is such a normal and strong human thing to do.
In fact, when we don't discuss China, most find modern software patents and copyright to be abusive.
Finally, copying UX standards has been done for ages, it gives users a comfortable way to start using a product and instead of fighting that to be unique, some rather embrace it.
It can be jarring using so many different apps if each one tries too hard to be unique.
And this is my two cents, here it makes sense to have a similar interface, it's ultimately an app with chat sessions.
I think you misunderstood. “Narrative” typically frames something as polemic or open to interpretation, which this is about equally so to the whole “spherical Earth” story.
> Kimi Work is a Local Agent designed for deep workflows. It mounts your local folders, navigates the web autonomously via WebBridge, runs Python code in the background, and executes scheduled tasks.
It's clearly a dupe of Claude/Codex products (Codex especially, styling-wise), but I think Kimi's goal here is simply to appear on feature-parity with bigger labs. I doubt they'd want to invest much into designing a UI for "the future of agentic work" from the ground up.
Their bread and butter and claim-to-fame will continue to be low-cost, near-frontier models, for which they need to at least appear to have some app layer to plug into.
And, in some ways, duping is a great way to show people that even the app layer can be commoditized.
I think we need to stop using the term "dupe" as if there was some originality somewhere that was not built upon or copied what already existed.
Some paradigms are simply ergonomic. Sometimes there's only one UX people will accept.
No one would buy a car steered with pedals; even if that is better (hey planes do it!) - every car maker is "copying" steering wheels and putting them on their cars. Is it controversial? If not why not?
> duping is a great way to show people that even the app layer can be commoditized
Everyone is already cloning the app layer. There are 30+ claude-code clone, and codex work already cloned cowork, grok and gemini is doing the same.
I highly recommend to anyone that uses them a lot - try to clone your own. I’ve got mine fully formed into how I want it, with features I haven’t seen anywhere else yet and with integration into everything I work with. We’ve lived in a world where every app had to fit enough people well enough - but now we live in a world where an app can be molded to be absolutely perfect for just you!
Out of curiosity, what are your favourite features that are unique to your implementation?
This is the way. Guess what am I working right now? My own “Work” app that will be better than anything else available today!
Same here! https://github.com/rush86999/atom
Looks like the thing I am building but of course mine will be slightly different. It would be great if we could all meet to show off and discuss our AI stuff.
How did you build your own coding agent? What language/framework did you use?
This comment is a great example of how large and strange the skills gap in AI is right now.
Curious why your first impulse is not simply to point your favorite agent at a few examples and start brainstorming/planning from there?
Multiple times I’ve built a purpose specific bespoke tool starting this way. In fact, it’s a great way to learn how specialized tools are built.
The problem of the sota models clamping down on third party harnesses should be stated upfront. Getting a SOTA model in a custom harness requires API pricing or risking an account ban, correct?
This preliminary change in cost may not make it worth it for people depending on which is their "favorite agent". Especially as the default harnesses for said agents continue to improve...
> Getting a SOTA model in a custom harness requires API pricing or risking an account ban, correct?
No, only Anthropic has that policy (and I think even that is relaxed for an unknown period if you use the Claude Agent SDK: https://support.claude.com/en/articles/15036540-use-the-clau...).
OpenAI, Kimi, Qwen, GLM and Deepseek all allow it.
I'm not sure about Gemini.
Presumably people want to hear the opinions of other people and still believe that they might learn more from the their experience than by exclusively interacting with AI models?
The alternative of that first impulse should be what exactly? Telling AI to research Medium AI drivel articles (since nobody seems to be posting long/medium form content anywhere else these days) to figure out what’s the best approach to building an AI harness?
BTW just my personal observation but I generally find AI notoriously bad at building any AI applications compared to standard software (which is not surprising given the scarcity of high quality training data)
It's pretty simple nowadays if you know conceptually how they work. Running the LLM calls in a loop with tools is an agent. You only need 10 or so basic tools to accomplish nearly anything, and you can build a dynamic skill system from that. Look at https://github.com/patw/pengy, ignore the app look at the spec.md file, feed that to your current agent of choice and make your own version. Use whatever tech stack or UI you're comfortable with. Change some of the choices in how it works, so it fits what you want to work.
Why would the language matter? Just use your favorite.
Look at pi coding agent.
It would have been much easier to clone six months ago but today Codex/Claude desktop apps are so feature packed it seems it wouldn’t be worth it.
For example, the mobile/remote control feature I find very useful but not sure it would be worth cloning. If you aren’t replicating every feature you aren’t exactly “cloning” it, you’ll just end up with a crappier version with a fraction of the features of the real thing.
But dont you have to be faithful to rl environment in which it model was trained in the harness?
only labs knows what the rl enviroment was so they will always be one step ahead.
> It's clearly a dupe of Claude/Codex products
Have you seen Kimi Code ?
It's hard to believe it's a different code than Claude Code.
In that case is there not some open harness that demand can coalesce around?
There are but then you're required to pay API prices for them which can end up costing significant multiples of the price of the subscription.
I am confused why they don't just offer a subscription with bring-your-own-harness, that would be a differentiator from claude/codex I think.
Actually unless they’ve changed substantially Kimi kind of does this? It’s not as advertised/focused on and the API is a little nonstandard making it hard to plug some things into, but, unlike Claude they’ll actually give you an API key you can use directly as part of your unmetered subscription. I played with it for a month or two out of curiosity, it ls not a bad value though a bit brittle and stuff
I think that pretty much all Chinese providers do that?
And Codex obviously allows that as well. Anthropic is pretty the main exception.
The Chinese api prices are already so cheap that that is arguably not needed.
You can also get $10/month opencode go (which officially can be used with any harness) with basically every major open weight model running inference in a place that I deem more trustworthy than the subject-to-Chinese-law inference options (but research this for yourself, there are some asterisks). This is what I use (in opencode) when I want an agent for personal use. Even if I'm being fairly intense with my usage by my standards anyway), I haven't managed to run out (yet), although I usually use the cheaper models (deepseek v4 is my favorite, and I've used both flash and pro).
From the FAQ:
> The plan is designed primarily for international users, with models hosted in the US, EU, and Singapore for stable global access. Our providers follow a zero-retention policy and do not use your data for model training.
And:
> Yes, you can use Go with any agent. Follow the setup instructions in your preferred coding agent.
Isn’t OpenCode just reselling Deepseek’s tokens over the same API? For v4-pro the prices from other providers on OpenRouter are way higher so what else could it be?
Opencode Go already seems extremely unprofitable if people use up at least 30% of their monthly tokens so its hard to imagine they’d go with several times more expensive inference provider yet show Deepseek pricing on their page.
Negative margins to have you hooked on someone else’s software?
I am gently presuming that they are not selling tokens at a loss in their all-you-can-eat plans like the frontier models are, but maybe I'm wrong.
Yup they have a cost advantage.
They just need to be good enough and just copy.
Sorry american labs - it’s the beginning of the end.
American labs will do fine with monthly subscriptions costing $20-300/month. They will not do fine for heavy basic LLM API calls -- it is for this that the cheap API produces substantial savings. Specialized niche provisions, e.g. excellent tool use, reliable structured output, reliable long context, cached and uncached web search, document reads, computer use, integrations, etc. also will take time for the foreign services to be good at.
So much is invested in the huge American AI companies that China undercutting them with clones that work almost as well can actually hurt the US economy.
They're at the point where copying can get people 90% of what they want anyway, the moat is drying up.
If you remove the word "AI" from your post it also describes the last 20 or so years of US/Chinese economic competition. They are real, serious competitor but also there is a lot of value in that last 10%. Some buyers need the best and you can charge a premium for it.
The speed in which a copier can catch up to a leader is faster than every before, but everything is faster these days. I think the threat is if can these labs catch up and pass American labs while keeping costs low, but I don't see that happening.
>Some buyers need the best and you can charge a premium for it.
Yes, and that is how you get a top-heavy k-shaped economy I believe. It is harder and harder to sell premium A+ products, because most people can only afford what works. If my old ice cream vendor charges me 25% more but puts chocolate sprinkles on it i'm going for the non-sprinkle version.
> the last 20 or so years of US/Chinese economic competition
Only with some industries, mostly electronics and light manufacturing. The rest of the history of software products doesn't look like this at all. China cloned Amazon and Meta fairly well (but only within their local/captured markets) but not Microsoft or Google and absolutely not Apple. Entertainment and culture exports are still effectively zero (like, what, three blockbuster films?) where the US, Japan and Korea define pop across the globe. Heavy industry is mixed, with decent success in consumer vehicles now but little purchase in stuff like heavy equipment or shipbuilding.
Basically it's complicated, and not really a point well suited to that kind of generalism.
> China cloned Amazon and Meta fairly well (but only within their local/captured markets)
Americans order Chinese products from SHEIN and Aliexpress type websites.
So China did penetrate the US ecommerce market, but they are not taking on Amazon directly.
People are buying from the Chinese and reselling on Amazon, which has a good logistics network that Chinese companies can't possibly build in the USA.
> absolutely not Apple
You need to look up Oneplus to see how many people were using their devices till they announced exit from Western markets.
The Chinese Android phones are essentially replacement for a big chunk of Apple ecosystem.
If these cheap mobile devices didn't exist - Apple could possibly be selling many more devices.
> China cloned .... Meta fairly well
TikTok is giving Meta a run for their money in USA.
They had to be restricted via threat of regulation to offload stake in TikTok USA.
> Microsoft
They didn't compete with Microsoft directly as entering Western Enterprise market is hard for the Chinese. Even Google can't beat Microsoft there till now.
“You need to look up Oneplus to see how many people were using their devices till they announced exit from Western markets.”
The final part of this sentence shows why there is absolutely no need to look up Oneplus.
Not British, but taking UK out of the definition of pop, is a bit too rough a cut. Also France. Look at the queues outside pop-glamorous stores. And was pop-art a pop artefact? Is TikTok redefining pop? Movies is an old measure, and even US has fallen with just franchise dominating. I think as a globe we are much more integrated and co-dependent, occasionally also in bad ways, than we appreciate.
I'm talking about economic activity more than the passports of the individual artists. And no, TikTok is a software product (a social media platform), not popular culture. There's absolutely popular culture being created and driven there but it's 100% NOT coming out of China. Are you high, as it were?
TikTok competes with Meta, not Olivia Rodrigo or BTS.
I am not high, I have kids, and can tell you they barely care about Olivia Rodrigo or BTS. You dispel passports as irrelevant for one leg but base your argument on them for the other. The contradiction is sufficient.
> Entertainment and culture exports are still effectively zero (like, what, three blockbuster films?) where the US, Japan and Korea define pop across the globe.
And this is really sad, because they have thousands of years of cultural heritage and also plenty of great films. Somehow, they've failed at exporting this completely.
> Heavy industry is mixed, with decent success in consumer vehicles now but little purchase in stuff like heavy equipment or shipbuilding
Uhh, China does over half of the world’s shipbuilding.
China is also responsible for 25% [1] of the world’s heavy equipment manufacturing and is considered to have double[2] the CAGR of the global average for heavy equipment manufacturing.
0: https://www.csis.org/analysis/china-dominates-shipbuilding-i...
1: https://www.grandviewresearch.com/horizon/outlook/constructi...
2: https://www.grandviewresearch.com/industry-analysis/heavy-co...
Xiaomi did a quite good job copying (and surpassing!) Apple.
They even got the car Apple always wanted, but never managed to build.
the difference is that the US economy is currently over-investing in AI to the point where it's a significant portion of the economy, it's the biggest competitive opportunity in quite a while
China isn't undercutting anything. Kimi has to be hosted somewhere and is extraordinarily expensive to operate at Fable-levels of capabilities. The moat is (and will always be) the cost to actually operate these models, and that cost will continue to soar, it will not reverse, and the hardware will never keep up such that at-home use somehow becomes economical.
It's like releasing a cutting edge search engine, proclaimed to be on par with Google. Oh yeah, I forgot to mention, you need to actually run it somewhere. One of the core moats for Google has been the cost to scale and compete globally with a search engine, which only a few entities on earth could even attempt. Somehow the HN crowd forgets this every single time. The same claims were endlessly made regarding Uber, as though the app was the hardest part of competing with Uber. Here's a hint: the model isn't the only super hard part of the total process at this point. Try running something akin to GPT 5.6 for a billion users, see how much that costs.
The winners will be the owners of the subscriptions and API usage. That's going to be OpenAI and Anthropic as of now, possibly Google will stay in the fight with their scale advantage focused on the consumer. It will never be the open weight models out of China. Who is going to invest the trillion dollars into infrastructure needed to run the latest version of Kimi 3-5 years out, for a billion users? There's only one entity that would ever consider it, and that's China domestic.
And the real money (for a sustainable business model allowing for constant massive reinvestment & infrastructure maintenance) is in pairing up advertising, which is another spot where the open weight models from China will never compete. It doesn't matter if the HN crowd hates this outcome, nothing will stop it from being the future core of mass consumer usage. And the business use will stay with subscription + API use as well: the best option for that today is OpenAI and Anthropic.
>Who is going to invest the trillion dollars into infrastructure needed to run the latest version of Kimi 3-5 years out
Everybody that is concerned about their data ownership and residency, and how the US political system is behaving right now. There is a huge incentive to cut off ties with API and subscription providers, especially those based in the US.
If you think Open AI can make hundreds of billions of dollars selling subscriptions for a chatbot I have a bridge you may be interested in.
> Their bread and butter and claim-to-fame will continue to be low-cost, near-frontier models, ...
K3 is beating Fable on several benchmarks. So I take it you don't consider Fable 5 to be frontier.
What are the frontier models?
People endlessly claimed the same thing about DeepSeek upon its release. they claim it every single time a very large cutting edge model is released out of China. It's a combination of propaganda, some truth, and the faux open culture desperation to see big tech get destroyed (somehow better if big China takes its place instead).
https://arena.ai/leaderboard/code
(that's just one, there are others... and several are claiming that those saying "yes sure it's just as good, but it uses more tokens" ain't really true either).
And I don't think anyone is excited about China releasing anything: I think people like the idea of open-weights.
People saying copy aren't wrong...but if you can offer a copy at 1/5th of the price then you've got a winning product not a copy
All bikes look the same, all cars look the same, all cameras look the same, all bridges, all dams, hell all smartphones look exactly the same! The whole goddamn point of UX is that you make something that is usable, and hence familiar to your users.
People who say "I was making GPUs before them, and they built their own GPUs that behave like GPUs so they copied me" don't understand how stuff is built. The hard part in engineering is not "imagining the high-level concept", it's actually making it work. Even if you can open electronics and study it to design and build your own, it requires a lot of expertise. And if you manage to do it, then you are an expert, period.
I'm sure many people find comfort in saying stuff like "You know what? I was a bit of a visionary because when I saw the first personal computer, I imagined that it would be cool if it could be orders of magnitudes faster and smaller such that it fits in a pocket". As if that was the hard part.
Typically, too many people spend a lot of time talking about UI and stuff like "unpacking" as if that was the most important part. But usually that's because they have no idea about how it works under the hood, so the only thing to discuss is whether the round edges are great or not.
A good UI/UX is noticed and appreciated by many. A great UI/UX goes completely unnoticed by most.
This is false.
Apple’s UI/UX is the reason why they are adored from a physical attraction perspective.
Apple UI/UX sucks though, but all the other alternatives also suck.
I think it can safely be argued that (only) good UX but also great marketing are the reason of that.
They have good UI, it's noticed and appreciated by many.
Yes, Apple's UI/UX is good, not great. Like far from great. So he's still correct
Is Kimi 1/5th the price? Artificial analysis has it at the same price as GPT 5.6 Sol while performing worse.
Yeah, but maybe Kimi doesn't pretend to know better than I do. Like, I just had a task today where I had a silly form where I wanted to copy in a signature from one document to another document. And, you know, I could have just used some other dumb tools, but I decided, hey, let me us ask Codex to do it. But it started lecturing me on how we can't do this kind of stuff. And it was just pure silliness. So hopefully Kimi Work doesn't try to pretend that it knows better and lets me decide on ethics.
This blog post here summed it up the best. These are tools, so let them be tools.
https://geohot.github.io/blog/jekyll/update/2026/07/11/ai-20...
computer says no
Winning product that can’t answer a simple question regarding Uyghurs or Tiananmen Square… it is a vulnerable product…
As opposed to Claude which has flagged:
- anything related to piracy
- legit security work (audit my infra, pentest it)
- disabling a safety mechanism on my food processor
- reverse engineering
- creating style guides inspired by other games/companies (for fan projects)
- questions about how nerve agents work and how they can be treated (for a layman)
I'd prefer it uncensored too but this objection was insightful in 2023. Everyone and their dog knows they're censored on topics sensitive to their country, much like western models are censored for western woke sensitivities. Nobody cares anymore as long as the reasoning is good and the price is low
Yes, “Nobody cares” is what scares me.
If we don't care about the other person than who we are?
On the other side, I am not a woke optimist that everything on Earth is lit by Sun…
It's an open model, abliterated versions will appear on HF in the fullness of time.
kimi has been particularly shameless in copying codex 1:1, wow.
Like, I get it, before condex there was conductor and a ton of other guis but this looks like they just copied the code.
Also, their privacy disclosure is incredibly misleading:
> How does Kimi Work protect my privacy when accessing local files? You have absolute control over your files. The built-in Ask before acting safeguard means Kimi will prompt you for explicit authorization before it modifies, overwrites, or runs code within your local directories. Nothing happens without your consent.
They fail to mention that is has unfettered read access to your files. ie, without your consent.
You have almost no control with Codex or Claude in that you have to trust them not to break their own rules.
You have almost no control with Kimi either, they can also say one thing and do another. You will not be able to sue a Chinese company if it breaks the law.
I can think of no faster way to hand over your trade secrets or intellectual property than to run something like this.
People will use Kimi anyway though because they are naive and don't value those things.
>They fail to mention that is has unfettered read access to your files.
Isn't that totally expected of a coding agent? Otherwise it's like complaining dropbox "has unfettered access to your files".
Yes, but the answer is an indirection.
“How does Kimi Work protect my privacy when accessing local files?”
It doesn’t protect your privacy. It’s like asking, “how do I know you’re not spying on me?” And getting the reply “we cannot physically enter your house.”
>It doesn’t protect your privacy. It’s like asking, “how do I know you’re not spying on me?” And getting the reply “we cannot physically enter your house.”
In light of the recent grok build snafu, "Nothing happens without your consent" seems like an upgrade. Moreover what are they supposed to do here? By that logic should any sort of non e2e email/storage provider not be able to put "How does [product] protect my privacy" in their faq?
Looking through their privacy policy, it looks like the bigger issue is that they are (or refuse to refute) training on your data/prompts. That's probably what people should be complaining about rather than complaining about how it "has unfettered access to your files", when that describes all coding harnesses.
Yes I’m saying a question answered deceptively should not be one of the four featured questions on your product page.
> Looking through their privacy policy, it looks like the bigger issue is that they are (or refuse to refute) training on your data/prompts.
You’re making my entire point. They’ve designed an ingestion engine that is not private by design. Any sort of privacy posturing is wrong.
"How does Kimi Work protect my privacy" can be interpreted in two ways. The first is "How does Kimi Work protect my privacy from Moonshot?" They did not answer this question. The second is "How does Kimi Work protect my privacy from other third parties?" They are answering this question by saying, we won't upload your files to some third party server without your consent.
It is, but I really think this project would be better as open source. They are already all-in on open for the most valuable part (the frontier model)! They might as well make the harness fully transparent so that questions about what's being accessed and transmitted can be avoided.
I personally wouldn't use any agentic harness from any company, American, Chinese, or otherwise, that is closed source. The Grok fiasco shows why.
The repo you’re working on, sure. But all files?
> The repo you’re working on, sure. But all files?
Yes, any tool, including any coding agent, has access to all files $USER has access to.
Sandbox it explicitly to give access to only the current directory - https://github.com/ashishb/amazing-sandbox
Sorry, I meant that’s not desirable behavior that the user might expect. I understand that coding agent tools tend to play fast and loose with access because it’s not their own privacy/security at stake.
This is true for all CLI tools on Mac and Linux (and other operating systems from that era).
Historically, everything you ran was trustworthy.
Android and iOS were invented in an era that does not allow this because the risks became evident by then.
Not sure what that's about, I was just copying the wording from OP. Their harness probably has the same restrictions as most other coding harnesses. That is, the read tool gives all project files by default without prompting, and everything else requires manual approval (or opting into auto-run all). If you decide to run it from your home directory however, then your "project" is all your files.
Everyone’s saying that it looks like Codex but TBH it just looks like all the AI chat websites on the side and some frequent actions on the top.
Sharing among AI friends what’s the problem, scraping the Internet or hiring new employees between companies who like to share? It’s a tradition among the AI model makers…
>kimi has been particularly shameless in copying codex 1:1, wow.
Isn´t codex (edited) Apache 2.0 (thanks @Shank)
https://github.com/openai/codex ?
No, it's Apache 2. https://github.com/openai/codex/blob/main/LICENSE
I guess the issue is they didn't give attribution?
Do we still fall for the China = must be spying, US = wouldn't spy unless it's disclosed?
It's fair to say whatever software you run from any country of origin is hoovering up as much data about you that it can get away with.
If you haven't compiled open source, audited code and are running on that, all bets are already off.
> kimi has been particularly shameless in copying codex 1:1, wow.
OpenAI's pigeons come home to roost
I am more concerned about what happens to the file if I give permission.
Opensource and local are seeming like the only way.
if your model ever touches bash, it has unfettered acess to all your files.
Not if the harness applies an OS-level sandbox to the process, as Codex does.
https://learn.chatgpt.com/docs/sandboxing?surface=cli
Unfortunately it was possible in codex too, until recently.
https://github.com/openai/codex/issues/5237
No, you need to apply the sandbox to the codex process itself. Codex does not do this.
It gets tricky when you want to say “allow this particular tool to have a little more access.” Like allow the codex client to access the GPT API but not some random other site the LLM learned about. Or like letting git write to .git to commit changes but don’t let something install git hooks. Or allow downloading certain packages/versions from a site like PyPI, or allow certain operations on certain GitHub repos in a way that doesn’t line up with how you can configure an access token.
The Dutch have a saying: it's better to have stolen something well than to have invented something bad.
Words to live by.
How dare they copy the largest plagiarism machine ever created.
Pretty shameless copy, but also shows how easy it is to do so. OpenAI and Anthropic had first-mover advantage, but easy come, easy go.
Safe to say there is no moat.
You mean other than the zillion dollars of infrastructure you need to run these systems for a billion users. And the extreme cost to sustain that operation day to day. And the cost to upgrade the infrastructure forever.
Remind me which entities (govt or corporate) are going to subsidize that forever. Remind me which entities are going to front the cost just to scale up.
By that premise - wherein you forget everything other than the first ingredient (the model) - Google has no moat either (got $100 billion dollars lying around?). Nor does Uber (have tens of billions in your bank account to clone their business?).
I'm certain building an app (Uber), or building a search engine (Google), is the super duper hard part.
Slapping a GUI on Linux will kill Windows, any day now.
Hyperscalers would generally be expected to sell compute to anyone who is willing to pay assuming that supply/demand for datacenter hardware evens out at some point that should become true again.
Sure if you want to bake a cake you also need to invent the universe first…
As long open weight models get released; no moats are going to exist. Both windows and Linux are weak comparisons.
If money is the only moat, you'll see race to bottom till ROCE goes down to gsec returns.
wow -- "Let's take something off your plate" just copied word-for-word from the Claude Cowork UI.
And it is above the fold in their hero image!
Good. Anthropic, OpenAI, Google built their models by shameless training on works of millions of people without any compensation.
I wish they had some kind of US hosted option, NDR policy, or something like that. I love the price and power, but I have concerns around data sovereignty and leaking IP to an overseas company.
I never really understood this sentiment. Some foreign company / government cannot do anything to me. My own government, on the other hand...
As soon as the K3 weights are published on July 27th, there will be many US providers hosting the model. I realize model hosting doesn't really apply to kimi work specifically, but in terms of accessing K3, there will be US options likely in 7 days. Just look at OpenRouter to see the providers hosting K2.6 https://openrouter.ai/moonshotai/kimi-k2.6.
Looking at these pages reminds me of browsing jellybean components on DigiKey. Amazing for the consumer, but one does wonder: What are all these DotAI infra companies’ plans to justify any of their valuations if they’re just selling commodity compute on OpenRouter?
It'll most likely land on Bedrock when the K3 weights are released.
Will it? Bedrock doesnt have deepseek v4, GLM 5.2, Qwen 3.6, minimax m3.
Fireworks AI has all of those, and is cheaper than Bedrock.
Azure should have those.
it has kimi 2.5, which isn't to say this will show up but who knows
Same. No way I am uploading data to a PRC company. It's not just an IP issue ... there's a very high likelihood such companies will be restricted in the United States, even if they try "Chinawashing" stunts like Bytedance.
I wish they'd embed the Code UI as a tab, similar to Claude Code Desktop.
There's really no real reason for those tools to be separate and for them not to give me something a bit nicer than just a TUI in the terminal.
They already have a locally runnable web UI when you have Kimi (the CLI) installed: https://www.kimi.com/help/kimi-code/cli-getting-started
Though it's ever so slightly broken (the sub-agent tracking stop button does nothing, some toggles in the Settings are not toggleable and so on, really slows down with long conversations to the point where keyboard input lags by seconds, think 600+ turns).Personally, I still liked ZCode better and will probably write a blog post about my experience with K3 too, the harness might need some more work, but the model is really, really good.
Just wire in "teh cloud!!1" to your local files.
What could go wrong?
i mean don't mount your nudes to it?
Still a chat interface, so definitely not an "AI Desktop".
Also this was released about a month ago
https://tech.yahoo.com/ai/meta-ai/articles/moonshot-ais-kimi...
Aren’t there all chat interfaces? Same goes for Claude Work, ChatGPT Work, etc
That's precisely the problem.
People are building UI/UX for a paradigm that is so utterly new that they miss the mark entirely.
Everyone defaults to a big chat input because that's all they know. It's the expectation, so it's safer, faster, and cheaper to not deviate.
It's also wrong and bad for work.
> It's also wrong and bad for work.
Say's who? Why? I honestly don't get the problem. OpenAI just did what you said with the Mac app and were forced to roll it back because of the uproar. A chat box is simple, relatable, easy to use. K.I.S.S. is a thing
> Says who.
Says everyone who's not using this stuff. It's an empirical argument. Is anyone actually using this for work? How long has it been since Cowork was released?
It's such an inadequate tool for work it reeks of SF hubris to think you can just slap a chat input box onto a PC and replace 90% of white collar jobs.
A CLI interface has been perfect for my coding work, why is it bad?
it's pretty great for coding because the CLI provides the model with all the context it needs and with a way to verify the result of its actions via linting, compiler messages and program output
the rest of "work" doesn't have that luxury
Copying a block of multi-line text out of the agent doesn't work over ssh, for instance.
How would you fix it?
Models need richer context to do well. They also generally need a verifier, but better context is already an improvement. So to fix it, you need to rethink UIs to be agent-first.
I'm not saying Kimi Work and its peers are entirely useless. I'm just saying they're a hack that puts a pretty low upper limit on agent capabilities and speed.
I feel that Kimi’s overall design is quite similar to Codex, and I hope it can excel over Codex in certain functional aspects.
Is there an opensource version of such a tool. I would really like a desktop app and use different models instead of a CLI.
OpenCode has a desktop version, it's decent: https://opencode.ai/download
Try paseo.sh . It does need the CLIs (claude, opencode, pi etc) running but it is a great interface for desktop and mobile.
The most astonishing thing to me about the AI boom is that nobody really has managed to establish vendor lock-in. Switching is so easy.
It's like this in the early days of any technology. It reminds me of the early days of the internet before everything became hyper optimised algorithms.
It's like the late stages of a monopoly game - it's not fun for anyone except the main winner.
The Pricing page shows all plans are sold out.
They say that they have too much demand so are stopping new subscriptions while they scale up: https://x.com/Kimi_Moonshot/status/2078855608565207130
> Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritizing compute for current members. Existing subscribed users are not affected.
> We're adding capacity as fast as we can and will reopen new subscription spots in batches.
That might even be a beneficial move, creating (well, actual, not artificial) scarcity works well to draw interest usually. Curious how this one plays out.
Yea, they had so much demand for Kimi K3 that they hit the limits of their compute: https://news.ycombinator.com/item?id=48969291
First time I've ever seen something like this. Fascinating.
Gotta generate that FOMO
I wonder when they are going to start realizing that on-premises self-hosted is the solution over local
The Chinese marketing blitz is on! Ha ha
Guess where the models are gonna be hosted? Hint, not China.
Looks like this can use webBridge, run Python and shell code in the background. Cowork can take over application UI as well which is still a differentiator (and a higher security risk) unless kimi offers it elsewhere.
They already have a hosted version of openclaw
https://www.kimi.com › bot
so...GUI/closed source version of their @MoonshotAI/kimi-code TUI?
https://github.com/MoonshotAI/kimi-code
Is there a link anywhere for a docs site for this? Really curious to know what "deep data sources" they have set up for the Native Global Market Data feature.
Most docs are here https://www.kimi.com/code/docs/en/kimi-code-cli/customizatio...
Not sure about the feature you mentioned but one can see in their app that they use IMF, World Bank, SEC EDGAR, and Yahoo Finance.
while it's true the model layer is getting commoditized, there is absolutely no way we are migrating any serious enterprise workflows to Kimi work. i think chinese models will continue to have an adoption problem.
Am i the only one that is wary of connecting kimi to my desktop and filesystem. Very different beast than using its API via OpenRouter.
Could I use the browsing feature to search for land to buy? Or is there an easier way I’m overlooking?
Hoping to find something on a creek with a Mountain View and fairly remote.
If history is anything to go by, in the next week we'll see a distill of GPT 5.6 or Mythos that is not as powerful as either, but pokes its head just in front of Kimi K3.
Agents are quite capable for any work involving CLI (not TUI) applications.
Seems to work for me to manage my emails. I have Proton Mail bridge and `mbsync` dumping all mail to a local folder, with an inotify watch script to react to new emails and classify them as either needing my attention (ignore) or something that does need my attention (send me a notification).
I've been looking for a good CLI spreadsheet editor
I'm kinda confused given that they're from China. The Chinese Internet is very different from the western Internet. The web is way smaller, and a lot of useful content is behind app-based walled ecosystems such as Wechat, Douyin and Xiaohongshu. This makes web automation much less useful. And I don't see them marketing something like Wechat automation?
Has anyone cancelled their $200 ChatGPT subscription and started using Kimi? It'd be an interesting metric to measure. I for one have no motivation to use this model. Nonetheless, I'm always happy to see more competition.
There should be a benchmark re Tiananmen Square and Uyghurs among LLMs.
Grok answered, ChatGPT answered, Claude answered.
Chinese Z.ai replied “Content Security Warning: The input text data may contain inappropriate content“.
So it is nice that Kimi is cheaper but for what price?
I've been subbing to Claude for 2 years and somehow the questions of whatever happened in Tiananmen (I know already anyway) have never once came up, but Claude (especially Fable) is now famous for refusing to work on things people want. And even pushed back against me when I was doing asking to research some people famous (infamous) in Israel-Palestine conflict. I really don't get why these "chinese gotchas" are keep coming up in every thread, they are mostly entirely inconsequential to getting actual work stuff done while same can't be said about Fable or others. Once, Gemini refused to help me make an infographic about women in STEM because the specific woman I was doing research about has died a few years ago, other models had no such qualms. All these "gotchas" are quite arbitrary.
I can second this. I was using Fable for some research and fable said it wasn't going to do it and downgraded to opus. Switched to K3 and continued my work. And it worked really well for what I was doing. You can easily switch claude code to use K3, it took like 30 seconds.
So for your Tiananmen questions go to Fable, and for the Fable-nerfed subject matter go to Kimi (or GLM, as Hugging Face were forced to do when they were recently hacked and the nerfed US models refused to analyze it).
that looks exactly like the recent codex!
When they release Kimi Design and I'll consider moving over
They're running like 10 tokens per second, constantly timing out, and launching new products?
How about deploy some inference GPUs first
Read the privacy policy.
It rubs me the wrong way that they have to copy everything, the UI is almost exact copy of Codex. The entire mind set of just copy the best is holding them back.
It's not about copy the best, it's about reducing friction for someone coming from a competing product.
Hence why everyone uses the same play, stop etc "buttons" / signs on "mediaplayers" whether virtual or real. [1]
Unless your user really trusts you / doesn't have an alternative reducing friction is better than being subjectively better [2]
[1] triangle, square etc...
[2] bottom placement versus top placement of browserbar on mobile phones... From an ergonomics perspective way better but from a UX experience worse... Hence chrome changed top to bottom but was forced to change it back on android, while apple forced their users from top to bottom and didn't have to move it back
Give them time, they'll work out their (Chinese UI/UX) groove just like they're working out EVs in the Chinese market. Copy first, improve and adjust later.
I hope they don't bring Chinese UI/UX to it. Native Chinese apps are a mess from a western perspective.
> It rubs me the wrong way that they have to copy everything, the UI is almost exact copy of Codex
I'm sure OpenAI haven't copied or stolen anything without permission - oh wait a sec..
The US loves capitalism because it pRomOtEs cOmpEtiTion. Well, here's the competition you asked for, lads - suck it up.
I have a UI I developed from first principles about 10 years ago, I use it to replace my os, its like my personal do everything emacs os but not at all emacs, and its funny to see other Uis iteratively convergence on it as context finally gets its due.
What's your UI like?
Kimi Work apparently
If this is what pops this bubble, I'll take it!
Nobody actually producing hard goods, real world value and actual software are using these cute tools. Its completely worthless to people actually doing things.
It's worthless until it isn't.
Will be cool until the AI goes ape shit and starts deleting all your files, wipes out your corporate file share, but not before uploading it all online.
Not saying Kimi will do this, but it's perfectly reasonable to expect this to happen, because there is really nothing anyone can do to prevent it. No, this is not meant to go in an isolated VM on a Linux box in a cheap VPS somehwere. This is meant for regular home users and corporate users to install on their primary system where all their personal and corporate data is available.
The above is a PSA (public service announcement) not just for Kimi, but also for GPT and Claude. If you don't understand how it's impossible to prevent the above from happening, then you should never use one of these tools.
No linux? They should ask k3 to add linux feature lol.
Not available for linux...
thats pretty disappointing.
China copying shit blatantly haha but winning surely
For the past 3 years, many software engineers were coping that AI (particularly LLMs) was only going after juniors. Given the documented history of mass layoffs related to AI from 2024 to today, it is clear that everyone including seniors and staff level folks are just as affected.
So let's just say, if your whole identity is for being a typical software engineer no matter the seniority and nothing else, perhaps it is time for you to reconsider options as of course 100% avoiding AI is futile at this point.
There will still be new jobs, but there will be no need for more excess people, even under Jevon's paradox. Software engineers were just the test bed for this and now there is an oversupply of SWEs without jobs in the market and will have to use AI tools like Kimi Work and Claude Cowork to be employable.
I have never been more sure about the viable future of software engineering as a career than I am after experiencing everything all the way up to and including Fable.
Do you actually believe this and are concerned about it?
First of all to reply to your deleted comment, yes this is completely relevant to the article.
As token costs drop significantly (thanks to local models and hosted Chinese models.) employers will consider using cheaper models like Kimi Work and others, even if they are already using expensive ones like Claude Cowork. They won't be able to operate without it.
Secondly, do I actually believe this? Of course, I already said that there will be more mass layoffs because of AI years ago. [0]
So this outcome was expected.
[0] https://news.ycombinator.com/item?id=42490692
Last year, not years ago
this landing page looks like a cheap ripoff of google's antigravity: https://antigravity.google/
Who cares. Competition is good for the user even if it's built on plagiarism. The other site was probably designed using an LLM anyway. LLMs can in some aspects converge wrt their outputs.
Also known as the "Xi Jinpeng learns about your business" app.
I surely hope so!
It's a mistake to make this such an obvious copy. It will feed the narrative that Chinese entities cannot be trusted by the West. They could have varied it even slightly. It's one thing with hardware because customers don't need to trust suppliers any further than they can throw the box. But these labs seem to be battling for enterprise knowledge work-share. Trust is paramount, and they're already behind on that front. Own goal.
Copying is not frowned upon in the LLM ecosystem. This is its very reason of existence.
How to make an apple pie from scratch with no copying:
Step 1: invent the universe
Funnily enough, you will have to pretty accurately copy the laws of our universe to make an Apple pie. Even small changes will change the properties of atoms, and hence all of chemistry and biology. And there won't be any apples or wheat.
The baker is not asking you to trust them with your entire knowledge base in exchange for an apple pie.
Why do you trust Sam Altman?
I don't trust him or Dario frankly, but I don't run a Fortune 500 company. They obviously do.
Perhaps, but "we run the exact same business as Anthropic and OpenAI to the pixel at 0.1% the price" is going to raise eyebrows.
Eyebrows will raise at Anthropic / OpenAI, perhaps. Why are they charging 1000x more for the exact same business?
The 'distillation is theft!' shtick won't last long. In fact, unlike Anthropic/OpenAI, these Chinese labs actually had to pay for that 'stolen' training data.
> Why are they charging 1000x more for the exact same business?
The implication they want to avoid, I'm suggesting, is because the CCP isn't plugged into and funding those entities' activities at a loss. Or they're lying about ZDR, etc. etc.
GLM 5.2 and Kimi 3 both had huge API pricing jumps (GLM most expensive chinese model, by a lot, then then Kimi 3 a lot higher than that). the cost advantage is rapidly decaying. the oft-cited cost per task makes Kimi look good vs Anthropic but even, or slightly vs OpenAI
Once the OpenAI/Anthropic-driven hype dies down people with decision-making power will be taking a look at the numbers and look at real productivity metrics. And they'll see it makes little sense to pay so much when competition is around and similarly good.
This is already happening.
Firms aren’t going to wait - their valuations are driven by it. So too is the compensation of those running them.
Why?
For fear that the data they're entrusted with will be copied or that there is some nefarious reason like state sponsorship that makes the activity economical?
Meh, we should just ignore that kind of propaganda. We copy things too, we use a system to protect IP because copying is such a normal and strong human thing to do.
In fact, when we don't discuss China, most find modern software patents and copyright to be abusive.
Finally, copying UX standards has been done for ages, it gives users a comfortable way to start using a product and instead of fighting that to be unique, some rather embrace it.
It can be jarring using so many different apps if each one tries too hard to be unique.
And this is my two cents, here it makes sense to have a similar interface, it's ultimately an app with chat sessions.
I do agree with the UX idea of familiarity, but copying the exact salutations, etc. just seems egregious.
Why not let the model speak for itself and contribute to an open harness?
Narrative?
Kind of like the sphere Earth narrative.
Pedantic.
> A narrative is a structured account of connected events, experiences, or ideas, commonly referred to as a story
Regardless of intent, it is how you tell your story.
I don't think the comment was intended to be pedantic, but rather to reframe the message as scientific fact.
I think you misunderstood. “Narrative” typically frames something as polemic or open to interpretation, which this is about equally so to the whole “spherical Earth” story.