Sable targets proofs of outputs from specified quantized models

Sable is an argument system for a specific claim about language-model output: that a committed, integer-quantized transformer produced a stated token sequence from a given prompt, bit for bit. The system defines a fixed integer function, F_O, with no floating-point step. According to the abstract, rounding, normalization, softmax and sampling decisions are determined by the inputs. That definition provides the computation the argument addresses; the abstract alone does not provide enough detail to inspect the full specification.

This scope matters. Sable’s claim concerns the output of a particular committed model under a defined computation. It is not a general proof that an answer is correct, nor a statement about every implementation or model.

How the argument is constructed

The abstract says Sable encodes the computation in a constraint system whose relations provably never wrap modulo the 31-bit KoalaBear prime. It combines zerochecks and LogUp-GKR lookups into one sumcheck per domain. It also commits to every witness tensor and model weight in one multilinear commitment, built from Reed–Solomon codes, BLAKE3 Merkle trees and a mixed-length BaseFold opening. These technical ingredients are reported in the abstract; without the paper’s full methods, their implementation and interactions cannot be examined here. The prover streams the model layer by layer on one GPU.

Reported performance

On one NVIDIA RTX PRO 6000 Blackwell GPU, the abstract reports proofs covering all 64 layers of Qwen2.5-32B in 92.5 seconds for 4,096 tokens and 398 seconds for 16,384 tokens. For 4,096 tokens, it gives 121.8 seconds for Qwen3-32B and 84.7 seconds for the mixture-of-experts Qwen3-30B-A3B. A verifier with a compiled core reportedly checks a proof on one CPU core in 19 to 56 seconds.

Proof generation is considerably slower than serving: the abstract reports a cost of 41 to 50 times serving time. It estimates 6.4 cents per 4,096-token request at public GPU prices. It also reports 40 to 64 times less time per parameter and token than zkLLM. These are the abstract’s reported figures; the evidence available here does not include the experimental details needed to assess the comparison methods, workloads or cost assumptions.

The abstract also compares Sable’s integer function with BF16 model outputs on selected evaluations. On WikiText-2, its greedy token matches the BF16 model’s at 86.3% of positions. On Qwen3, the abstract reports results within 1.6 percentage points of BF16 on MMLU and 1.7 points on GSM8K. These figures apply to the named evaluations; the abstract does not establish broader equivalence between the computations.

Security claim and limits of the evidence

The abstract reports round-by-round soundness error of 2^-103 in the random-oracle model. This is a formal bound stated under that model. The abstract does not provide the full security analysis, so the assumptions and proof details cannot be assessed from the material available here.

Sable is not zero-knowledge, and the abstract says it targets open-weight models. That limitation is important: the reported system is not presented as a privacy-preserving proof. The abstract does not give enough information to characterize more specifically what a proof may disclose.

This account relies on the arXiv abstract, not the full methods, results tables or appendices. The source is a preprint; the material available here does not establish its peer-review status. The abstract links to code, but that link alone does not independently verify the implementation or reported measurements. Sable’s abstract describes a way to argue for a narrowly defined model-output claim, alongside substantial proof-generation costs and evaluations whose details require the full paper.