QuarterBit – Train 70B LLMs on a single GPU

3 comments

• quarterbit 11 hours ago

  I built QuarterBit because AI training costs are insane. A 70B model needs 840GB of memory — that's 11 A100 GPUs at $30+/hour.

  QuarterBit AXIOM compresses training memory 15x. Same model. Same quality. Fraction of the hardware.

  RESULTS:

    Llama 70B: 840GB → 53GB (11 GPUs → 1 GPU) = 90% savings
    Llama 13B: 156GB → 9GB (FREE on Kaggle T4) = 100% savings
  

  91% energy reduction vs standard training. 100% trainable weights (not LoRA/adapters). 3 lines of code.

  HOW IT WORKS:

    from quarterbit import axiom
    model = axiom(model)
    model.cuda()
  

  TRY IT:

    pip install quarterbit
  

  Demo (FREE): https://www.kaggle.com/code/kyleclouthier/quarterbit-axiom-1...

  Benchmarks: https://quarterbit.dev

  AXIOM uses a novel weight representation combining lossless compression with a built-in optimizer. Weights stored at 0.62 bytes/param vs 4 bytes FP32. Gradient updates happen directly in compressed space.

  Not quantization-aware training or LoRA — every parameter fully trainable, convergence matches AdamW.

  Solo founder from Canada. Self-taught CUDA/ML. Applying to YC S26.

  Happy to answer questions.

  [-]
  • smallerize 11 hours ago

    Put two spaces at the beginning of a line for monospace.

      Like this

    [-]
    • quarterbit 11 hours ago

      Thanks for the tip! It won't let me edit it. I think I will hide and repost it does look sloppy.