A new system for proving large model outputs are exact
Sable provides a way to prove that a specific model produced a given sequence of tokens bit for bit.
Researchers created Sable, an argument system and single-GPU prover. It proves a committed, integer-quantized transformer produced a specific token sequence from a prompt. This process is bit for bit exact.
The system uses an integer function with no floating-point step. Every rounding, normalization, softmax, and sampling decision is determined by its inputs. The relations in the constraint system never wrap modulo the 31-bit KoalaBear prime.
Sable batches zerochecks and LogUp-GKR lookups into one sumcheck per domain. It commits to every witness tensor and weight using one multilinear commitment. This commitment is built from Reed-Solomon codes, BLAKE3 Merkle trees, and a mixed-length BaseFold opening.
The prover streams the model layer by layer on one GPU. On one NVIDIA RTX PRO 6000 Blackwell GPU, Sable proves all 64 layers of Qwen2.5-32B in 92.5 s for 4,096 tokens. It also proves Qwen3-30B-A3B in 84.7 s for 4,096 tokens.
A verifier with a compiled core checks a proof on one CPU core in 19 to 56 s. The proof is not zero-knowledge. This means the system targets open-weight models.
Key facts
- Sable proves all 64 layers of Qwen2.5-32B in 92.5 s for 4,096 tokens on one NVIDIA RTX PRO 6000 Blackwell GPU.
- The proof costs 41 to 50 times the serving time, which is 6.4 cents per 4,096-token request at public GPU prices.
- In the random-oracle model, the interactive proof has a round-by-round soundness error of .