Sable Provides Bit-Exact Verifiable Inference for Quantized Transformers
November 9, 2026
Sable is a single-GPU prover that provides bit-exact proofs for integer-quantized transformer outputs. It uses a constraint system based on a 31-bit KoalaBear prime and streams model layers to prove token sequences from prompts.
HOW THIS AFFECTS YOU
●
builderYou can verify that a quantized model is producing the exact outputs it claims to produce.
●
researcherThis allows for mathematical proofs of inference correctness in quantized systems.