Language model proofs take 41 to 50 times serving time

Solberg reports single-GPU cryptographic proofs for 32-billion-parameter models, but the guarantee applies to a precisely defined integer version, not the original floating-point model.

Industry

Tomer Solberg

MoonMath.ai

Research Digest·12 November 2026·2 min read

Illustration: Language model proofs take 41 to 50 times serving time

In one line

Sable proves, on a single GPU, that a committed integer-quantized language model produced a specified token sequence exactly.

Before you read this

Three ideas this paper builds on

verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.327 papers →large language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.306 papers →post-training quantizationQuantizing pre-trained neural network weights and activations to50 papers →

Read it as

Customers receiving text from a language model usually cannot verify which model settings or calculations produced it. Computer rounding can vary, so a trustworthy check needs a precisely defined version of every calculation.

Sable defines that version using whole-number arithmetic with fixed rounding rules, then creates a cryptographic proof: evidence others can check without repeating the calculation. A single graphics processor builds the proof one model layer at a time.

It proves a large model’s output in 92.5 seconds on one graphics processor. That makes independent audits more practical, but proving takes much longer than producing the text, and this precisely defined version can choose different words from the standard model’s approximate decimal arithmetic. It is designed for auditing models whose settings are public.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

What it released

Code

Concepts

verification sumcheck argument large language model post-training quantization multilinear polynomial commitment mixture of experts