Back to Tool
F

FlashProve

Tool Nov 3, 2026

Anthropic's open research artifact for cryptographic proofs that a model response matches committed weights, without publishing those weights.

Visit primary link 2 sources attached
Anthropic announcement illustration showing metallic chain links over four paper backgrounds.
Original cover illustration for Anthropic’s November 2 provable-inference research announcement. This is an announcement visual, not a benchmark chart or evidence of deployed model verification. Source: Anthropic
self-hosted

What happened lately

FlashProve timeline

FlashProve separates a provider’s commitment to model weights, the production of inference proofs and a client’s verification of those proofs. Its released research implementation targets Llama 4 Maverick with a short context window and integer arithmetic under greedy decoding.

The repository is Apache 2.0 licensed and describes itself as a research artifact that is not maintained and does not accept contributions. Its prover requires a supported Linux and NVIDIA H100 stack; checking a proof uses a CPU. A successful model-identity check does not establish that the response is factually correct.