# Tomer Solberg presents Sable, a proof system for bit-exact language-model inference [Canonical reading](https://escapemeridian.com/readings/tomer-solberg-presents-sable-a-proof-system-for-bit-exact-3c83589a) Tomer Solberg presents Sable, an argument system and single-GPU prover that produces a proof that an integer-quantized transformer generated a particular token sequence from a prompt, bit for bit. Sable defines an integer function with no floating-point step, then uses a constraint system designed not to wrap modulo the 31-bit KoalaBear prime. On one NVIDIA RTX PRO 6000 Blackwell GPU, it proves all 64 layers of Qwen2.5-32B in 92.5 seconds for 4,096 tokens, and in 398 seconds for 16,384 tokens. The paper reports that making a proof costs 41 to 50 times the serving time; a verifier checks one on a CPU core in 19 to 56 seconds. Solberg reports that the function’s greedy token matches the BF16 model’s at 86.3% of WikiText-2 positions, and that it is within 1.6 points on MMLU and 1.7 points on GSM8K for Qwen3. The proof is not zero-knowledge, so Solberg says Sable targets open-weight models. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Nov 10, 2026 (2026-11-10). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-11T17:36:38.179Z. First observed: 2026-11-11T05:06:46.350Z. Retrieved: 2026-11-11T17:34:35.966Z. - [Sable: A Small-Field Sumcheck Argument for Bit-Exact Inference of Large Language Models](https://arxiv.org/abs/2611.14083) ## Limits The abstract reports selected models, workloads, and evaluation results; it does not establish performance or token agreement for models and workloads beyond those described. The proof is not zero-knowledge, and the paper therefore targets open-weight models. ## Open question How would Sable’s reported proving costs and token-match results change across models and workloads beyond those tested in the paper? ## Source-stated Solberg presents Sable as a system that proves an integer-quantized transformer generated a given token sequence from a prompt, bit for bit. > We present Sable, an argument system and single-GPU prover for the statement that a committed, integer-quantized transformer produced a given token sequence from a given prompt, bit for bit. - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source-stated The paper reports proving all 64 layers of Qwen2.5-32B on one NVIDIA RTX PRO 6000 Blackwell GPU in 92.5 seconds for 4,096 tokens and 398 seconds for 16,384 tokens. > 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 and 398 s for 16,384 tokens - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source-stated The paper reports proof costs of 41 to 50 times serving time, while a verifier checks a proof on one CPU core in 19 to 56 seconds. > A proof costs 41 to 50 times the serving time, or 6.4 cents per 4,096-token request at public GPU prices - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source-stated On the reported evaluations, Sable’s function has a greedy token match with the BF16 model at 86.3% of WikiText-2 positions and is within 1.6 and 1.7 points on two Qwen3 evaluations. > On WikiText-2, F\_O's greedy token matches the BF16 model's at 86.3% of positions; on Qwen3, F\_O is within 1.6 points of BF16 on MMLU and 1.7 points on GSM8K. - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source-stated Because the proof is not zero-knowledge, Solberg says the system targets open-weight models. > The proof is not zero-knowledge, so the system targets open-weight models. - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source and contributor context Solberg, Tomer · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Model inference Using a trained model to produce a prediction or response to new input. - [Definition reference](https://developers.google.com/machine-learning/glossary#inference) ## Graphics processing unit A chip that performs many calculations at once; the main hardware used to train and run AI models. - [Definition reference](https://en.wikipedia.org/wiki/Graphics_processing_unit) ## Quantization Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy. - [Definition reference](https://developers.google.com/machine-learning/glossary#quantization) ## CPU The central processing unit: a general-purpose processor that carries out a computer's instructions. - [Definition reference](https://aws.amazon.com/what-is/cpu/) ## Open weights A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted. - [Definition reference](https://opensource.org/ai/open-weights) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## NVIDIA NVIDIA develops processors, networking and software used for AI, graphics and scientific computing. - [Explore this context](/entities/nvidia) - [Reference](https://blogs.nvidia.com/) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 063d8752-7fb3-403f-89d8-f68f0fc0ba58. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T17:36:38.179Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: a1f46d299060f2a30b1bb85f97be6f15e422b6cbd09aad5293aec01827c4c756. Source text hash: dca2a90499e1fc28b1fd056e625e8c931691456c67031ded9c7076ae0bb5b3a5. ## Revision and source-check history ### Version 1 — 2026-11-11T17:36:38.179Z Tomer Solberg presents Sable, an argument system and single-GPU prover that produces a proof that an integer-quantized transformer generated a particular token sequence from a prompt, bit for bit. Sable defines an integer function with no floating-point step, then uses a constraint system designed not to wrap modulo the 31-bit KoalaBear prime. On one NVIDIA RTX PRO 6000 Blackwell GPU, it proves all 64 layers of Qwen2.5-32B in 92.5 seconds for 4,096 tokens, and in 398 seconds for 16,384 tokens. The paper reports that making a proof costs 41 to 50 times the serving time; a verifier checks one on a CPU core in 19 to 56 seconds. Solberg reports that the function’s greedy token matches the BF16 model’s at 86.3% of WikiText-2 positions, and that it is within 1.6 points on MMLU and 1.7 points on GSM8K for Qwen3. The proof is not zero-knowledge, so Solberg says Sable targets open-weight models. ## Original source arXiv cs.AI · Article Original source date: Nov 10, 2026 (2026-11-10). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-11T17:36:38.179Z. First observed: 2026-11-11T05:06:46.350Z. Retrieved: 2026-11-11T17:34:35.966Z. - [Sable: A Small-Field Sumcheck Argument for Bit-Exact Inference of Large Language Models](https://arxiv.org/abs/2611.14083) ## Limits The abstract reports selected models, workloads, and evaluation results; it does not establish performance or token agreement for models and workloads beyond those described. The proof is not zero-knowledge, and the paper therefore targets open-weight models. ## Open question How would Sable’s reported proving costs and token-match results change across models and workloads beyond those tested in the paper? ## Source-stated Solberg presents Sable as a system that proves an integer-quantized transformer generated a given token sequence from a prompt, bit for bit. > We present Sable, an argument system and single-GPU prover for the statement that a committed, integer-quantized transformer produced a given token sequence from a given prompt, bit for bit. - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source-stated The paper reports proving all 64 layers of Qwen2.5-32B on one NVIDIA RTX PRO 6000 Blackwell GPU in 92.5 seconds for 4,096 tokens and 398 seconds for 16,384 tokens. > 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 and 398 s for 16,384 tokens - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source-stated The paper reports proof costs of 41 to 50 times serving time, while a verifier checks a proof on one CPU core in 19 to 56 seconds. > A proof costs 41 to 50 times the serving time, or 6.4 cents per 4,096-token request at public GPU prices - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source-stated On the reported evaluations, Sable’s function has a greedy token match with the BF16 model at 86.3% of WikiText-2 positions and is within 1.6 and 1.7 points on two Qwen3 evaluations. > On WikiText-2, F\_O's greedy token matches the BF16 model's at 86.3% of positions; on Qwen3, F\_O is within 1.6 points of BF16 on MMLU and 1.7 points on GSM8K. - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source-stated Because the proof is not zero-knowledge, Solberg says the system targets open-weight models. > The proof is not zero-knowledge, so the system targets open-weight models. - [Supporting source](https://arxiv.org/abs/2611.14083) ## Source and contributor context Solberg, Tomer · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Model inference Using a trained model to produce a prediction or response to new input. - [Definition reference](https://developers.google.com/machine-learning/glossary#inference) ## Graphics processing unit A chip that performs many calculations at once; the main hardware used to train and run AI models. - [Definition reference](https://en.wikipedia.org/wiki/Graphics_processing_unit) ## Quantization Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy. - [Definition reference](https://developers.google.com/machine-learning/glossary#quantization) ## CPU The central processing unit: a general-purpose processor that carries out a computer's instructions. - [Definition reference](https://aws.amazon.com/what-is/cpu/) ## Open weights A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted. - [Definition reference](https://opensource.org/ai/open-weights) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## NVIDIA NVIDIA develops processors, networking and software used for AI, graphics and scientific computing. - [Explore this context](/entities/nvidia) - [Reference](https://blogs.nvidia.com/) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 063d8752-7fb3-403f-89d8-f68f0fc0ba58. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T17:36:38.179Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: a1f46d299060f2a30b1bb85f97be6f15e422b6cbd09aad5293aec01827c4c756. Source text hash: dca2a90499e1fc28b1fd056e625e8c931691456c67031ded9c7076ae0bb5b3a5. ### uncertain — 2026-11-11T17:36:38.179Z — anthropic/claude-opus-5.5 Publication: 063d8752-7fb3-403f-89d8-f68f0fc0ba58 attribution — pass: The paper is correctly attributed to its sole listed author, Tomer Solberg. Results are framed as presented or reported by the paper rather than as established facts. > Authors:Tomer Solberg summary — uncertain: The figures match the abstract, but a general reader likely cannot tell what Sable is for. Central terms are left unexplained: argument system/prover, integer-quantized transformer, bit for bit, floating-point, modulo the KoalaBear prime, zero-knowledge, open-weight, BF16, greedy token, MMLU and GSM8K. There is also no plain statement of the stake the abstract supports: a checkable proof that a specific committed open-weight model produced a given output. Smallest repair: add one sentence in plain language on what the proof lets someone check, and tooltip or gloss the key terms. Minor: 'designed not to wrap' softens the source's 'provably never wrap'. > a committed, integer-quantized transformer produced a given token sequence from a given prompt, bit for bit date — pass: The originalPublished date of 2026-11-10 matches the arXiv v1 submission date. The feed date of 11 November is correctly not used. > \[Submitted on 10 Nov 2026\] support — pass: Each claim is stated in the abstract. This includes the proving times, the 41 to 50 times serving cost, the 19 to 56 second verification time on one CPU core, the 86.3% token match, and the MMLU and GSM8K gaps. The open-weight limitation is also stated. The question about generalizing beyond the tested models is a fair open issue rather than a leading claim, and no unsupported benefits are implied. > A verifier with a compiled core checks a proof on one CPU core in 19 to 56 s. media — pass: The selected media is the verified arXiv abstract page of this paper, which gives the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/3c83589af35c4864a2e51d778bbd8252ce888c2017d925c1f0a8e8e38eec5ac0/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/tomer-solberg-presents-sable-a-proof-system-for-bit-exact-3c83589a.json)