developer tools
The Model Output Gets a Bit-Exact Proof
Sable proves that an integer-quantized transformer produced a specific sequence, layer by layer, on one GPU.
Summary
Sable proves that an integer-quantized transformer produced a specific sequence, layer by layer, on one GPU.
Sable defines every rounding, normalization, softmax and sampling decision as an integer function and proves the resulting inference trace. The authors report proving 4,096-token runs of 32-billion-parameter models in 84.7 to 121.8 seconds on one RTX PRO 6000 GPU. The proof is not zero-knowledge and is aimed at open-weight models.
Why it matters
Sable proves that an integer-quantized transformer produced a specific sequence, layer by layer, on one GPU.
Limits and context
- The proof is not zero-knowledge and is aimed at open-weight models.
Key claims
Sable proves that an integer-quantized transformer produced a specific sequence, layer by layer, on one GPU.
Qualification: The proof is not zero-knowledge and is aimed at open-weight models.
Evidence: source-2026-11-11-020
Sources
- arXiv:2611.14083v1arXiv / Cornell University · primary research
Corrections
No corrections have been recorded for this story.