This logic seems mad. If people only need SLMs then hyperscalers can also centrally host higher-efficiency models, and still gain efficiencies of scale and convenience over hosting locally.
You have to remember that articles like these are written for finance people who don't understand the underlying technology, by finance people who don't understand the underlying technology. In this case, the author is a "CFA Institute Enterprising Investor", previously a CIO and basically their entire life been "money, money & money", so hardly surprising they're pulling a lot of assumptions based on what they read.
Read the paper the author talks about yourself instead (https://arxiv.org/abs/2511.07885), and also, contrary to what the author says in the article, do not do investments based on single papers made from academic studies, regardless of how much money this guy tells you you can make.
The valuations of the hyperscalars won't sustain just being more efficient than something you can run locally. There's a market there, but it's for margin on a commodity. They're priced for oligopoly on unique, premium products.
I very much don’t want to run it locally. I want the same one running somewhere else that I can interact with from all my devices. Look at something like Grok Bot. Nobody is going to run this locally. You can already self host almost anything, yet most people and businesses don’t.
But it is quite likely a 1T market. That won't support the stock valuations of OpenAI, Anthropic or Nvidia, but it will still mean massive data center build out and little relief to RAM prices.
Exactly the type of thing I’d pay someone not to deal with. It’s like media streaming vs piracy. Even though I could get a nice local setup with plex and have higher quality streams with free pirated br rips, I just don’t want to deal with it.
Maybe I could use codex or whatever to set that all up nowadays, but it was such a pain doing the initial setup and getting it all running and finding sources of media yadda yadda like a decade ago when I last tried lol
> Exactly the type of thing I’d pay someone not to deal with.
Very much a personality and/or lifestyle thing, but I've also been building infra for decades. I mean it was two container startups and a phone app download (10min). Not exactly difficult. If I didn't like it I certainly wouldn't be in tech!
> pay someone
I'd do that with something like 3 story roof work or foundation work with a contractor, but tech? If I want the outcome I will be happy with, often have to do it myself.
Yes, it's all about the amount of money invested right now. Justifying that was always going to be tricky, and it still is.
The compute capacity, however, is here to stay. There will be a seemingly endless line of people queued up to buy it at pennies on the dollar if this whole thing blows up financially. And they will use it!
And any data center with reasonable energy costs and without massive interest payments will still be profitable at pennies on the dollar. Might need some bankruptcies of the shell corporations to get rid of interest payments, but they will be very profitable post bankruptcy.
Hyperscalers don't run computing at some multiple more efficient than on prem. The only way hyperscalers can compete is if they own the sand to cycles supply chain and ensure that raw compute is out priced in the market (ram,flash,compute). Ram and flash were an easy target because they are a commodity in name only.
> Hyperscalers don't run computing at some multiple more efficient than on prem.
I'd disagree here. I see two avenues for an efficiency multiple, albeit a single-digit multiple:
* Client aggregation allows a hyperscaler to average out demand spikes from uncorrelated clients, reducing the peak:average demand ratio and allowing better budgeting of compute.
* Dynamic batching allows typical requests to run in batches of more-than-1 and/or overlap, offering better internal compute utilization ratios (e.g. interleaving output and input streams). The small limit of on-device LLMs will run with batch sizes of one with strong memory bandwidth bottlenecks.
For an example of these factors in action, see the API cost differential between batch, standard, and 'fast' processing. OpenAI prices these tiers at a 1:2:4 ratio.
1. Training a large model with lots of information, then stripping the "useless" information from that model to obtain a small model => nobody has shown this.
2. Training a small model, letting it use a database tool so it scores the same as a large model without database => nobody has shown this.
I partly agree. Theory (of computation) shows it must be possible, however nobody has produced the small model (well, depends who you ask, what is small, article disputes that) and the database yet.
To go very small (thousands of rules) so that the reasoner can be understood by humans and proven sound - might be computationally quite difficult.
This really elegantly sums up two of the most major areas of research in the field right now. I expect people are going to hammer at both problems for a while. It will be interesting to see how much useful stuff is built as a result.
I mean maybe yes? The hypothesis from the early GPT days was (and in a small way still remains): "If we just chuck more data into the training, does it get better at X, even if the data was seemingly unrelated to X?", and the workings of LLMs seem to kind be pointing in that direction, although with some ceiling.
But seemingly models good at programming for example, would get worse at programming if you removed everything not-programming. Train a model solely on syntax, and it'll be worse than a general purpose LLM on syntax, in general at least.
"But seemingly models good at programming for example, would get worse at programming if you removed everything not-programming. Train a model solely on syntax, and it'll be worse than a general purpose LLM on syntax, in general at least."
That might be because emergence of capabilities to reason about programs requires abstractions (such as fuzzy and modal logic) that are rarely present in software sources. That doesn't mean the reasoning model itself has to be large; neither does it have to emerge from the ML training on large language corpus, we might construct it by different means.
Sure. I don't give a damn about them. But Phil Collins man. I love reviewing his body of work before I get worked up. Obvously we cannot cut him, because I need him. So how do you decide what to leave out?
The cost you pay is in additional reasoning the SLMs have to do. As I write elsewhere, LLM "remembers" that "Socrates is mortal", or other commonly useful deduction. SLM might need to derive it first by reasoning from the DB, which slows it down. (Or worse, it might miss the correct reasoning because it's just too much side quests to follow.) But the advantage is flexibility.
That sounds a little bit like the guy who said that we don't need a computer in the home.
There may be a great deal of advantage to be able to run a small language model on something I have in my hand, disconnected.
Although mainframes have their use, the pendulum of centralized to distributed has gone back and forth and there are benefits to be gleaned from either model, sometimes at the same time.
From what’s presented this seems to be the lower end of Q&A and reasoning tasks and not long horizon agentic work. I agree that the search engine replacement AI usage is something that can run anywhere (though it’s still better run in the cloud for speed, context length, sandboxing and convenience) but this isn’t the engine of AI growth.
Also, the average consumer is not going to be running a local model until they are built into the hardware they already buy and when they are, who is supplying the weights? They’ll likely be shipped as an ASIC (or MSIC) at that point anyways. Those will use a licensed model from the current leaders. The whole argument sounds like saying that cloud services shouldn’t be profitable because everyone has a computer at home or to meme “we have AI at home”.
> Also, the average consumer is not going to be running a local model until they are built into the hardware they already buy and when they are, who is supplying the weights?
Why not? Apple would have you download an updated set of weights with OSX / iOS updates. Nvidia could bundle them with driver updates. I don't think either would struggle to get people to keep their weights up to date.
> Those will use a licensed model from the current leaders.
Says who? In the world of video decoding, H.265 is losing out to AV1 largely because it's not superior enough to H.264 to justify the expensive and complicated licensing.
Do you really think chipmakers are going to pay a 25c per unit tax to get a model that is 15% better?
I don't think they are toast, I mean they will be in some trouble because all of them have fallen victim to fomo and started building out with so much debt for capacity that may or may not be needed nor achieve the returns that they want.
I think there's a future where "personal software" meaning highly custom apps generated by an agent is a thing that doesn't mean everything will become that, same for local LLMs but all of this is still too far. The main issue is that hyperscalers or big tech in general have become too powerful they can just buy their way in and out of legislation as they please, sorry I mean lobby ... funny how if you rename something it becomes legal or illegal
Exactly. Seems naive at best for an investment consultant to look narrowly at current model capability and not consider the broader market. For example:
1. The hyperscalers are in a positive reinforcement loop. Despite any suggestions to the contrary they keep getting bigger. And can, er, “influence” government policy/officials and anything else needed to keep it that way.
2. The frontier labs and their investors. Another self-fulfilling reinforcement loop. Witness the circular gymnastics among OAI/Anthropic, Microsoft/Amazon and Nvidia
3. Data. No-one believes that Zuckerberg and co are going to say “great, we can just run the models on devices we don’t own and stop the surveillance economy because, y’know, privacy matters and we really care about mental health”.
And then there’s data centre locations and “yeah but jobs” even though your power bills are going up, and “why run your own data centre Mrs CTO, let us do it for you and save all that capex and those pesky employees you need to do it”.
Don’t get me wrong: I’m rooting for local, open weight/source models. But “hey look they benchmark well” is an unhelpfully narrow basis to forecast the demise of central hyperscaler hosting.
How many developers here don't see a difference between the latest LLMs and SLMs they can run on their own computer? I tried running a smaller model locally, and it's not usable for me.
I know people like to "predict" things, so that if they happen they can then say "I am a visionary, I predicted it" and start their blog posts with "as I predicted long ago (because I am a visionary), ...".
> The research report estimates that the addressable market in the US for SLMs has grown to about $10tn or one-third of the entire US GDP of $30tn. There isn’t much left for LLMs to thrive in, and every year, their advantage over SLMs is shrinking.
I stopped counting the number of times "estimates" said that a market would absolutely explode, and it absolutely didn't. Those are in the business of being a broken clock.
If something better comes, it will be better. Sure. And we would like to have something better, because it would be better.
> I tried running a smaller model locally, and it's not usable for me.
If you have the hardware, a MacBook Pro for Qwen 3.6 35B A3B and Gemma 4 26B A4B for example, they are absolutely usable, both in terms of speed and quality. Anecdotally, I can use Qwen for day-to-day coding tasks in TS and Go, without hickups.
I’ve been experimenting with Qwen 3.8 27B and I believe I can totally use it as my main coding model provided I have the hardware for the full context. I don’t need my model to be opus level. I need it to do the tasks I want it to do without being an overprotective nanny.
I'm unable to find a local model that comes close to the effectiveness of GPT models in Codex, and I have 96GB of VRAM available and tried every local model under the sun so far. Neither of those you mention I'd say are good enough for day to day software engineering for me, but I'm also really strict about code quality and iterate on what outputs agents give me a lot before I'm happy.
With local models, this iteration cycle takes maybe 30 minutes for a single fix or feature, rather than 10 minutes with GPT+Codex, as there is so many corrections and iterations needed, although I will say that the speed I'm able to get locally makes it more fun that any of the remote models.
> I'm unable to find a local model that comes close to the effectiveness of GPT models in Codex
I agree with you there, the local models are not as capable as frontier remote LLMs. If you‘re used to letting fable run for an hour to do a novel or complex task, you‘re (probably?) not going to be happy with local LLMs. But, to me, for my daily work, they‘re still usable, and often surprisingly capable. YMMV
> although I will say that the speed I'm able to get locally makes it more fun that any of the remote models.
This is becoming increasingly important to me. Super smart max reasoning frontier is fine if I leave it running overnight on some prepared set of clearly defined tasks, but when I want to work with the LLM, throughput really matters, and I'll go with a dumber model to get there.
At some point though, it's fast enough and any speed gains beyond that just makes me the bottleneck.
I also am seeing the smaller models gaining big strides lately, closing the gap on frontier models (still a decent sized gap though). I don't even run the small models like Qwen 3.8 27B locally. I just try them out in the cloud to see how they are progressing, and I'm definitely able to be productive.
No, you cant. I challenge anyone who claims this to show me an actual project built only by SLM's and not using opus, sonnet, sol, or terra. Spoiler: you can't.
>> I stopped counting the number of times "estimates" said that a market would absolutely explode, and it absolutely didn't. Those are in the business of being a broken clock.
The lack of logic and risk management on this statement, is so strong, I hope humans are all quickly substituted by LLMs. Lets just do it and be done with it...
If this is correct I see a future where the hyperscalers are funded by the businesses integrating siloed SLMs in their software.
Also the defence/intelligence industry will always want to keep an edge so don't be surprised if they stick around and we see favourable regulations for them similarly to how the government turns a blind eye to social media platforms because they increase the footprint of mass surveillance.
I wouldn't be surprised if the hyperscalers became software auditors and any piece of critical software was required to have a regulated security audit before it could enter production. Selling the poison and the cure is a great business model.
The paper underlying this blog post is fundamentally flawed because of benchmark ceilings. If we define only simple tasks like asking what is the capital of France, all models will converge to 100%, obviously. But as bigger models get more capable we want them to replace more and more complex tasks, in as short time as possible.
Then of course there is the economics of it. Do people prefer to spend $5000 upfront to get things done 5x slower, or would they rather pay $20 a month for that?
I believe it is true, and likely there exists a class of even smaller models than what they call "small".
You can imagine a reasoning model as a huge set of rules that generate the next statement from previous statements (written in context). In that sense, a reasoning model can be compared to a logical theory - you have certain deduction rules which can generate new judgments.
Often, logical theories are structured that the rules are remade into axioms, and the deduction rule is only modus ponens (which corresponds to function application and is a building block of program execution).
In the case of an LLM, the set of rules (or axioms) they have in the theory is quite large, but most likely semantically unsound (with respect to their their own representation of truth) - that's why LLM's make mistakes.
It would be desirable to break the logical theory represented by LLM into a smaller set of axioms, which would:
a) remove rules easily deductible from the smaller core of axioms (for example, LLM doesn't need to remember "Socrates is mortal", as it can derive it from "Socrates is a man" and "all men are mortal")
b) remove rules that have low value (facts that aren't used often or have weak validity) which cause ruleset to become unsound
I suspect that's what SLM distillation is doing, to some extent.
The question is, how far this process can go? I personally believe there is a useful logic for commonsense reasoning that has less than thousand rules (still several orders more than your typical mathematical logic, but orders less than SLMs). These axioms do not contain much facts about the world, but that could be added.
So I believe there is a sweet spot (deductive core, encyclopedic shell) which we have not yet found (it's a little bit more formal language than natural language) but is very efficient for general reasoning.
One of the big things to think about is whether local LLMs will be things companies want to deploy.
If you think of for e.g. some proprietary piece of software that wants to embed an LLM they've fine tuned or trained, they will want to make back some of their research cost right. So they are not going to want to put this on-device even if the hardware is there, unless there's some way of locking it down. I suspect we'll need on-hardware validation/verification and a way of preventing extraction of weights for this move to happen for many use cases.
A remaining advantage of large language models is that as they get larger, they tend to hallucinate less, simply because the odds of the training set containing a desired answer improve with size. If a solid "I don't know" detector is developed for inference, then you can try a small language model first.
An implication is that successful research in "I don't know" detection could destroy hundreds of billions in shareholder value.
Another "cool but we don't know how yet" thing would be a "confidence interval" so we know how much to trust LLM responses. Or while we're fantasizing, they could just know everything all the time regardless of training data. The "if a solid" part is easy to imagine, hard to implement :)
If you are like Google or Apple and you are delivering AI to a mass market unwilling to pay a lot for it, you are absolutely going to drive AI processing to endpoint devices. You are also going to spend what it takes in R&D make a hybrid system that knows when to use local compute or cloud compute. That's going to be the bulk of the workload.
The paper focuses on "intelligence per watt (IPW)", as a way to compare SLMs vs LLMs.
What might happen is that a chunk of the market, whatever its size will be, will end up going to SLMs run on iphones or Macbooks, and eat some of the revenues from LLMs, because not everyone needs the most powerful LLM all the time.
This is up to October 2025 though, yes? Obviously things are continually moving but Opus 4.5 launched in November and that was a recognised step change in capability. An up to date comparison would be interesting.
Given how often the "experts" on CNBC and Bloomberg TV use the term hyperscalers, my approach is to completely disregard anything written or spoken by any person who uses the term.
Two weeks ago, CNBC invited one of their "experts" who focus on SpaceX, and he said they have 10 million satellites in orbit. This is the current level of financial journalism available on "specialized" financial channels...
> they provide a better or at least as good an answer as LLMs in 62.5% of the cases.
Are we going to scrap hospitals because a vet could do the job 62.5% of the time?
The economics also point away from everyone buying a big RAM Mac that sits idle 99% of the time. SLM and own hardware sounds efficient and “free” but it is nothing of the sort when you factor everything in (and forfeit the sharing efficiencies of API)
SLMs are great esp for task specific fine tunes but this take isn’t it
> If their results are true, then we will hardly need any data centres in the future, and the hyperscalers are wasting hundreds of billions of dollars in investments.
What if they get sufficiently powered and watered industrial warehouses close to where the successful people live?
Cool, they scored well on all the "make complex calculations and I'll vibe check your results based on my own domain experience" things I use the average LLM chatbot for.
So maybe in 10yr I'll be able to run a SLM on a 5yo laptop and not have Google or whoever hoover up everything.
As much as I want local and open-weights models to succeed, nothing beats a paid frontier model for now. Anybody who claims otherwise is simply not a daily user of such models. So this "investor" here should invest sime time in actually using the various LLM models and get a real taste of what it's like.
Their point isn’t that local models are better or even as good more but that if you can do 50%+ of tasks with local then that’s 50% of tokens that aren’t captured as compute done in data centers.
> As you can see, on average, SLMs are as good if not better than LLMs in 81.2% of the cases, with the LLMs having a significant advantage only in areas like engineering, life sciences, transportation and computer sciences.
So what I’m reading here is “LLMs have a significant advantage” in the most critical areas that have practically infinite demand for more intelligence.
The article is kinda dumb, and yes this is clearly the area where frontier models having and advantage matters the most, but I'd point out that these smaller open-weight models are performing better than the big Frontier models of just 4-6 months ago.
This means that the Frontier labs are under immense pressure to maintain that lead, and could end up in serious trouble if they stumble at all.
The other thing id point out is that a lot of us who are token-sensitive do things like build plans using expensive, smart models, and then execute those plans using cheaper dumber models.
Then there's the fact that we are still in the age of heavily subsidized Frontier subscriptions + tokenmaxxing initiatives from megacorps. Neither of which are sustainable, and will drive more usage to smaller open models once they end.
While that's true, the open / local models are getting good enough. Given time and the technology trend people may prefer a private local model for most use cases. Nobody is arguing that a Ferrari isn't a faster car, but the Honda is the more practical choice.
You completely missed the thesis here, and that is supported by the numbers being presented. It is that a large share of ordinary inference can be routed away from the hyperscalers.
It is not the whole story, and knowledge is very lacking, but it has gotten a lot of attention. That model together with DeepSeek V4 Flash are the highlights of this summer on the open/local models side.
There is a likely US scenario and a rest of the world scenario. It will be interesting to see if China acts on the overextension of the US Military in the Middle East. Taiwan will be a big play for both and crucial to the hyperscalers.
But since the US is dabbling in piracy again and telling people what they can do and not do with their shit, it’s not too far-fetched that everyone that is not a global superpower is at risk of getting bombed to smithereens if they are a danger to US AI supremacy.
This is such a crazy timeline, predicting even like a single year ahead feels like looking into a medieval glass ball.
But we are humans, I am confident we will find a way to fuck this up royally for everyone. Brace for impact.
This logic seems mad. If people only need SLMs then hyperscalers can also centrally host higher-efficiency models, and still gain efficiencies of scale and convenience over hosting locally.
You have to remember that articles like these are written for finance people who don't understand the underlying technology, by finance people who don't understand the underlying technology. In this case, the author is a "CFA Institute Enterprising Investor", previously a CIO and basically their entire life been "money, money & money", so hardly surprising they're pulling a lot of assumptions based on what they read.
Read the paper the author talks about yourself instead (https://arxiv.org/abs/2511.07885), and also, contrary to what the author says in the article, do not do investments based on single papers made from academic studies, regardless of how much money this guy tells you you can make.
The valuations of the hyperscalars won't sustain just being more efficient than something you can run locally. There's a market there, but it's for margin on a commodity. They're priced for oligopoly on unique, premium products.
I very much don’t want to run it locally. I want the same one running somewhere else that I can interact with from all my devices. Look at something like Grok Bot. Nobody is going to run this locally. You can already self host almost anything, yet most people and businesses don’t.
Yes, but that's not a 10T business.
But it is quite likely a 1T market. That won't support the stock valuations of OpenAI, Anthropic or Nvidia, but it will still mean massive data center build out and little relief to RAM prices.
> I very much don’t want to run it locally. I want the same one running somewhere else that I can interact with from all my devices.
I run models on my desktop and access them from a phone app on the go. Wireguard tunnel.
Responds fast and lets me kick off tasks or workflows via text or voice.
Exactly the type of thing I’d pay someone not to deal with. It’s like media streaming vs piracy. Even though I could get a nice local setup with plex and have higher quality streams with free pirated br rips, I just don’t want to deal with it.
Maybe I could use codex or whatever to set that all up nowadays, but it was such a pain doing the initial setup and getting it all running and finding sources of media yadda yadda like a decade ago when I last tried lol
> Exactly the type of thing I’d pay someone not to deal with.
Very much a personality and/or lifestyle thing, but I've also been building infra for decades. I mean it was two container startups and a phone app download (10min). Not exactly difficult. If I didn't like it I certainly wouldn't be in tech!
> pay someone
I'd do that with something like 3 story roof work or foundation work with a contractor, but tech? If I want the outcome I will be happy with, often have to do it myself.
Yes, it's all about the amount of money invested right now. Justifying that was always going to be tricky, and it still is.
The compute capacity, however, is here to stay. There will be a seemingly endless line of people queued up to buy it at pennies on the dollar if this whole thing blows up financially. And they will use it!
Indeed, "pennies on the dollar" is the key point the author is making in this article.
And any data center with reasonable energy costs and without massive interest payments will still be profitable at pennies on the dollar. Might need some bankruptcies of the shell corporations to get rid of interest payments, but they will be very profitable post bankruptcy.
> And they will use it!
Just out of curiosity: Use it for what?
Whatever it is suited for that they can get a decent margin on. Or, you know, hobbies.
The demand is there and the market will find pricing that works.
Hyperscalers don't run computing at some multiple more efficient than on prem. The only way hyperscalers can compete is if they own the sand to cycles supply chain and ensure that raw compute is out priced in the market (ram,flash,compute). Ram and flash were an easy target because they are a commodity in name only.
> Hyperscalers don't run computing at some multiple more efficient than on prem.
I'd disagree here. I see two avenues for an efficiency multiple, albeit a single-digit multiple:
* Client aggregation allows a hyperscaler to average out demand spikes from uncorrelated clients, reducing the peak:average demand ratio and allowing better budgeting of compute.
* Dynamic batching allows typical requests to run in batches of more-than-1 and/or overlap, offering better internal compute utilization ratios (e.g. interleaving output and input streams). The small limit of on-device LLMs will run with batch sizes of one with strong memory bandwidth bottlenecks.
For an example of these factors in action, see the API cost differential between batch, standard, and 'fast' processing. OpenAI prices these tiers at a 1:2:4 ratio.
The logic seems mad to me because SLMs can simply not hold as much information as an LLM.
Maybe if you combine an SLM with a database (as a tool) then it could work, but someone should first prove that.
But do you really need a model that has the complete Duran Duran discography memorized and preloaded in RAM at all time?
That's a different question. Probably not. But:
1. Training a large model with lots of information, then stripping the "useless" information from that model to obtain a small model => nobody has shown this.
2. Training a small model, letting it use a database tool so it scores the same as a large model without database => nobody has shown this.
I partly agree. Theory (of computation) shows it must be possible, however nobody has produced the small model (well, depends who you ask, what is small, article disputes that) and the database yet.
To go very small (thousands of rules) so that the reasoner can be understood by humans and proven sound - might be computationally quite difficult.
This really elegantly sums up two of the most major areas of research in the field right now. I expect people are going to hammer at both problems for a while. It will be interesting to see how much useful stuff is built as a result.
But what if I _really like_ Duran Duran???
I mean maybe yes? The hypothesis from the early GPT days was (and in a small way still remains): "If we just chuck more data into the training, does it get better at X, even if the data was seemingly unrelated to X?", and the workings of LLMs seem to kind be pointing in that direction, although with some ceiling.
But seemingly models good at programming for example, would get worse at programming if you removed everything not-programming. Train a model solely on syntax, and it'll be worse than a general purpose LLM on syntax, in general at least.
"But seemingly models good at programming for example, would get worse at programming if you removed everything not-programming. Train a model solely on syntax, and it'll be worse than a general purpose LLM on syntax, in general at least."
That might be because emergence of capabilities to reason about programs requires abstractions (such as fuzzy and modal logic) that are rarely present in software sources. That doesn't mean the reasoning model itself has to be large; neither does it have to emerge from the ML training on large language corpus, we might construct it by different means.
Sure. I don't give a damn about them. But Phil Collins man. I love reviewing his body of work before I get worked up. Obvously we cannot cut him, because I need him. So how do you decide what to leave out?
The cost you pay is in additional reasoning the SLMs have to do. As I write elsewhere, LLM "remembers" that "Socrates is mortal", or other commonly useful deduction. SLM might need to derive it first by reasoning from the DB, which slows it down. (Or worse, it might miss the correct reasoning because it's just too much side quests to follow.) But the advantage is flexibility.
You mean like how Western Digital is only a 168 billion dollar business, but Dropbox is a $7.4B behemoth... er, wait?
That sounds a little bit like the guy who said that we don't need a computer in the home.
There may be a great deal of advantage to be able to run a small language model on something I have in my hand, disconnected.
Although mainframes have their use, the pendulum of centralized to distributed has gone back and forth and there are benefits to be gleaned from either model, sometimes at the same time.
From what’s presented this seems to be the lower end of Q&A and reasoning tasks and not long horizon agentic work. I agree that the search engine replacement AI usage is something that can run anywhere (though it’s still better run in the cloud for speed, context length, sandboxing and convenience) but this isn’t the engine of AI growth.
Also, the average consumer is not going to be running a local model until they are built into the hardware they already buy and when they are, who is supplying the weights? They’ll likely be shipped as an ASIC (or MSIC) at that point anyways. Those will use a licensed model from the current leaders. The whole argument sounds like saying that cloud services shouldn’t be profitable because everyone has a computer at home or to meme “we have AI at home”.
> Also, the average consumer is not going to be running a local model until they are built into the hardware they already buy and when they are, who is supplying the weights?
Apple or Nvidia, presumably.
Hardware yes, weights? lol
Why not? Apple would have you download an updated set of weights with OSX / iOS updates. Nvidia could bundle them with driver updates. I don't think either would struggle to get people to keep their weights up to date.
> Those will use a licensed model from the current leaders.
Says who? In the world of video decoding, H.265 is losing out to AV1 largely because it's not superior enough to H.264 to justify the expensive and complicated licensing.
Do you really think chipmakers are going to pay a 25c per unit tax to get a model that is 15% better?
I don't think they are toast, I mean they will be in some trouble because all of them have fallen victim to fomo and started building out with so much debt for capacity that may or may not be needed nor achieve the returns that they want. I think there's a future where "personal software" meaning highly custom apps generated by an agent is a thing that doesn't mean everything will become that, same for local LLMs but all of this is still too far. The main issue is that hyperscalers or big tech in general have become too powerful they can just buy their way in and out of legislation as they please, sorry I mean lobby ... funny how if you rename something it becomes legal or illegal
Exactly. Seems naive at best for an investment consultant to look narrowly at current model capability and not consider the broader market. For example:
1. The hyperscalers are in a positive reinforcement loop. Despite any suggestions to the contrary they keep getting bigger. And can, er, “influence” government policy/officials and anything else needed to keep it that way.
2. The frontier labs and their investors. Another self-fulfilling reinforcement loop. Witness the circular gymnastics among OAI/Anthropic, Microsoft/Amazon and Nvidia
3. Data. No-one believes that Zuckerberg and co are going to say “great, we can just run the models on devices we don’t own and stop the surveillance economy because, y’know, privacy matters and we really care about mental health”.
And then there’s data centre locations and “yeah but jobs” even though your power bills are going up, and “why run your own data centre Mrs CTO, let us do it for you and save all that capex and those pesky employees you need to do it”.
Don’t get me wrong: I’m rooting for local, open weight/source models. But “hey look they benchmark well” is an unhelpfully narrow basis to forecast the demise of central hyperscaler hosting.
"If", sure.
How many developers here don't see a difference between the latest LLMs and SLMs they can run on their own computer? I tried running a smaller model locally, and it's not usable for me.
I know people like to "predict" things, so that if they happen they can then say "I am a visionary, I predicted it" and start their blog posts with "as I predicted long ago (because I am a visionary), ...".
> The research report estimates that the addressable market in the US for SLMs has grown to about $10tn or one-third of the entire US GDP of $30tn. There isn’t much left for LLMs to thrive in, and every year, their advantage over SLMs is shrinking.
I stopped counting the number of times "estimates" said that a market would absolutely explode, and it absolutely didn't. Those are in the business of being a broken clock.
If something better comes, it will be better. Sure. And we would like to have something better, because it would be better.
> I tried running a smaller model locally, and it's not usable for me.
If you have the hardware, a MacBook Pro for Qwen 3.6 35B A3B and Gemma 4 26B A4B for example, they are absolutely usable, both in terms of speed and quality. Anecdotally, I can use Qwen for day-to-day coding tasks in TS and Go, without hickups.
I’ve been experimenting with Qwen 3.8 27B and I believe I can totally use it as my main coding model provided I have the hardware for the full context. I don’t need my model to be opus level. I need it to do the tasks I want it to do without being an overprotective nanny.
I'm unable to find a local model that comes close to the effectiveness of GPT models in Codex, and I have 96GB of VRAM available and tried every local model under the sun so far. Neither of those you mention I'd say are good enough for day to day software engineering for me, but I'm also really strict about code quality and iterate on what outputs agents give me a lot before I'm happy.
With local models, this iteration cycle takes maybe 30 minutes for a single fix or feature, rather than 10 minutes with GPT+Codex, as there is so many corrections and iterations needed, although I will say that the speed I'm able to get locally makes it more fun that any of the remote models.
> I'm unable to find a local model that comes close to the effectiveness of GPT models in Codex
I agree with you there, the local models are not as capable as frontier remote LLMs. If you‘re used to letting fable run for an hour to do a novel or complex task, you‘re (probably?) not going to be happy with local LLMs. But, to me, for my daily work, they‘re still usable, and often surprisingly capable. YMMV
> although I will say that the speed I'm able to get locally makes it more fun that any of the remote models.
This is becoming increasingly important to me. Super smart max reasoning frontier is fine if I leave it running overnight on some prepared set of clearly defined tasks, but when I want to work with the LLM, throughput really matters, and I'll go with a dumber model to get there.
At some point though, it's fast enough and any speed gains beyond that just makes me the bottleneck.
I also am seeing the smaller models gaining big strides lately, closing the gap on frontier models (still a decent sized gap though). I don't even run the small models like Qwen 3.8 27B locally. I just try them out in the cloud to see how they are progressing, and I'm definitely able to be productive.
No, you cant. I challenge anyone who claims this to show me an actual project built only by SLM's and not using opus, sonnet, sol, or terra. Spoiler: you can't.
You let a hiccup slip through in your comment though.
prooof-read by Qwen
>> I stopped counting the number of times "estimates" said that a market would absolutely explode, and it absolutely didn't. Those are in the business of being a broken clock.
The lack of logic and risk management on this statement, is so strong, I hope humans are all quickly substituted by LLMs. Lets just do it and be done with it...
> Lets just do it and be done with it...
Presumably not what you intended but this phrase immediately takes me to:
https://www.youtube.com/watch?v=dJFR7xbOIuw&t=42s
Great movie...yeah I think I was inspired by the scene... :-)
> I hope humans are all quickly substituted by LLMs
Why don't you go talk to your LLM instead of commenting here, then?
> I tried running a smaller model locally, and it's not usable for me.
Probably a skill issue on your part.
Haven't the SLMs been distilled using the LLMs?
If this is correct I see a future where the hyperscalers are funded by the businesses integrating siloed SLMs in their software.
Also the defence/intelligence industry will always want to keep an edge so don't be surprised if they stick around and we see favourable regulations for them similarly to how the government turns a blind eye to social media platforms because they increase the footprint of mass surveillance.
I wouldn't be surprised if the hyperscalers became software auditors and any piece of critical software was required to have a regulated security audit before it could enter production. Selling the poison and the cure is a great business model.
The paper underlying this blog post is fundamentally flawed because of benchmark ceilings. If we define only simple tasks like asking what is the capital of France, all models will converge to 100%, obviously. But as bigger models get more capable we want them to replace more and more complex tasks, in as short time as possible.
Then of course there is the economics of it. Do people prefer to spend $5000 upfront to get things done 5x slower, or would they rather pay $20 a month for that?
I believe it is true, and likely there exists a class of even smaller models than what they call "small".
You can imagine a reasoning model as a huge set of rules that generate the next statement from previous statements (written in context). In that sense, a reasoning model can be compared to a logical theory - you have certain deduction rules which can generate new judgments.
Often, logical theories are structured that the rules are remade into axioms, and the deduction rule is only modus ponens (which corresponds to function application and is a building block of program execution).
In the case of an LLM, the set of rules (or axioms) they have in the theory is quite large, but most likely semantically unsound (with respect to their their own representation of truth) - that's why LLM's make mistakes.
It would be desirable to break the logical theory represented by LLM into a smaller set of axioms, which would:
a) remove rules easily deductible from the smaller core of axioms (for example, LLM doesn't need to remember "Socrates is mortal", as it can derive it from "Socrates is a man" and "all men are mortal")
b) remove rules that have low value (facts that aren't used often or have weak validity) which cause ruleset to become unsound
I suspect that's what SLM distillation is doing, to some extent.
The question is, how far this process can go? I personally believe there is a useful logic for commonsense reasoning that has less than thousand rules (still several orders more than your typical mathematical logic, but orders less than SLMs). These axioms do not contain much facts about the world, but that could be added.
So I believe there is a sweet spot (deductive core, encyclopedic shell) which we have not yet found (it's a little bit more formal language than natural language) but is very efficient for general reasoning.
One of the big things to think about is whether local LLMs will be things companies want to deploy.
If you think of for e.g. some proprietary piece of software that wants to embed an LLM they've fine tuned or trained, they will want to make back some of their research cost right. So they are not going to want to put this on-device even if the hardware is there, unless there's some way of locking it down. I suspect we'll need on-hardware validation/verification and a way of preventing extraction of weights for this move to happen for many use cases.
> One of the big things to think about is whether local LLMs will be things companies want to deploy.
Non-tech enterprise was already doing this years ago.
Regulatory reasons, privacy reasons, security, etc. They want on-prem and total ownership of the data. Sometimes air-gapped.
A remaining advantage of large language models is that as they get larger, they tend to hallucinate less, simply because the odds of the training set containing a desired answer improve with size. If a solid "I don't know" detector is developed for inference, then you can try a small language model first.
An implication is that successful research in "I don't know" detection could destroy hundreds of billions in shareholder value.
>> A remaining advantage of large language models is that as they get larger, they tend to hallucinate less
First time I hear that...not really true.
"Understanding Why Language Models Hallucinate: Testing Reasoning Against Priors" - https://arxiv.org/abs/2607.00447
"Calibrated Language Models Must Hallucinate" - https://arxiv.org/abs/2311.14648
"TruthfulQA: Measuring How Models Mimic Human Falsehoods" - https://arxiv.org/abs/2109.07958
Another "cool but we don't know how yet" thing would be a "confidence interval" so we know how much to trust LLM responses. Or while we're fantasizing, they could just know everything all the time regardless of training data. The "if a solid" part is easy to imagine, hard to implement :)
If you are like Google or Apple and you are delivering AI to a mass market unwilling to pay a lot for it, you are absolutely going to drive AI processing to endpoint devices. You are also going to spend what it takes in R&D make a hybrid system that knows when to use local compute or cloud compute. That's going to be the bulk of the workload.
The paper focuses on "intelligence per watt (IPW)", as a way to compare SLMs vs LLMs.
What might happen is that a chunk of the market, whatever its size will be, will end up going to SLMs run on iphones or Macbooks, and eat some of the revenues from LLMs, because not everyone needs the most powerful LLM all the time.
This is up to October 2025 though, yes? Obviously things are continually moving but Opus 4.5 launched in November and that was a recognised step change in capability. An up to date comparison would be interesting.
Given how often the "experts" on CNBC and Bloomberg TV use the term hyperscalers, my approach is to completely disregard anything written or spoken by any person who uses the term.
Two weeks ago, CNBC invited one of their "experts" who focus on SpaceX, and he said they have 10 million satellites in orbit. This is the current level of financial journalism available on "specialized" financial channels...
What would be a better term?
It has to be the Hyperspenders
I think one of the watershed developments will be fast models.
Imagine current frontier models at 20k tokens/second.
There's also video models, world models, robotics simulations, the matrix...
Small language model gave satisfactory healthcare output in 100% of cases?
Complete nonsense.
> they provide a better or at least as good an answer as LLMs in 62.5% of the cases.
Are we going to scrap hospitals because a vet could do the job 62.5% of the time?
The economics also point away from everyone buying a big RAM Mac that sits idle 99% of the time. SLM and own hardware sounds efficient and “free” but it is nothing of the sort when you factor everything in (and forfeit the sharing efficiencies of API)
SLMs are great esp for task specific fine tunes but this take isn’t it
> If their results are true, then we will hardly need any data centres in the future, and the hyperscalers are wasting hundreds of billions of dollars in investments.
What if they get sufficiently powered and watered industrial warehouses close to where the successful people live?
Cool, they scored well on all the "make complex calculations and I'll vibe check your results based on my own domain experience" things I use the average LLM chatbot for.
So maybe in 10yr I'll be able to run a SLM on a 5yo laptop and not have Google or whoever hoover up everything.
As much as I want local and open-weights models to succeed, nothing beats a paid frontier model for now. Anybody who claims otherwise is simply not a daily user of such models. So this "investor" here should invest sime time in actually using the various LLM models and get a real taste of what it's like.
Their point isn’t that local models are better or even as good more but that if you can do 50%+ of tasks with local then that’s 50% of tokens that aren’t captured as compute done in data centers.
> As you can see, on average, SLMs are as good if not better than LLMs in 81.2% of the cases, with the LLMs having a significant advantage only in areas like engineering, life sciences, transportation and computer sciences.
So what I’m reading here is “LLMs have a significant advantage” in the most critical areas that have practically infinite demand for more intelligence.
The article is kinda dumb, and yes this is clearly the area where frontier models having and advantage matters the most, but I'd point out that these smaller open-weight models are performing better than the big Frontier models of just 4-6 months ago.
This means that the Frontier labs are under immense pressure to maintain that lead, and could end up in serious trouble if they stumble at all.
The other thing id point out is that a lot of us who are token-sensitive do things like build plans using expensive, smart models, and then execute those plans using cheaper dumber models.
Then there's the fact that we are still in the age of heavily subsidized Frontier subscriptions + tokenmaxxing initiatives from megacorps. Neither of which are sustainable, and will drive more usage to smaller open models once they end.
While that's true, the open / local models are getting good enough. Given time and the technology trend people may prefer a private local model for most use cases. Nobody is arguing that a Ferrari isn't a faster car, but the Honda is the more practical choice.
You completely missed the thesis here, and that is supported by the numbers being presented. It is that a large share of ordinary inference can be routed away from the hyperscalers.
How does a current local model compare to the best frontier model 12 months ago. Or 24 months ago?
It beats a frontier model from 12 months according to this bench: https://news.ycombinator.com/item?id=49334544
It is not the whole story, and knowledge is very lacking, but it has gotten a lot of attention. That model together with DeepSeek V4 Flash are the highlights of this summer on the open/local models side.
I have found qwen 3.8's coding quality using opencode to be similar to claude or gpt from 6-9 months ago, except much slower.
I think it’s a bit more complicated.
There is a likely US scenario and a rest of the world scenario. It will be interesting to see if China acts on the overextension of the US Military in the Middle East. Taiwan will be a big play for both and crucial to the hyperscalers.
But since the US is dabbling in piracy again and telling people what they can do and not do with their shit, it’s not too far-fetched that everyone that is not a global superpower is at risk of getting bombed to smithereens if they are a danger to US AI supremacy.
This is such a crazy timeline, predicting even like a single year ahead feels like looking into a medieval glass ball.
But we are humans, I am confident we will find a way to fuck this up royally for everyone. Brace for impact.