Anthropic releases FlashProve: cryptographic proof that a reply came from the committed model
According to Anthropic's research post (Nov 2, 2026), FlashProve is an open-source prototype of provable inference for large Mixture-of-Experts models. The provider publishes a commitment to the model weights and then produces a zkSNARK proof for each request; users verify the proof on an ordinary CPU without knowing the weights. For Llama 4 Maverick (a 400B-parameter MoE, 4-bit weights), a single H100 generates proofs at about 110 tokens/second, each proof about 19 MB for a 1,024-token request, and checking takes about 17 seconds on 24 CPU threads. Anthropic says GPU time is roughly 70–95x that of running vLLM, and the proof applies to an integer model under greedy decoding; production deployment has not been announced.
Source: Anthropic
Need an LLM chatbot or agent for your business?
Describe your need briefly; you will get clarifying questions and a scoped proposal before any work starts.
Related news
Tencent releases Hy4: first vision and computer-use support, aimed at the top tier of open models
LeanGuard: checking every tool call of an LLM agent against a policy written in Lean 4
