ResearchHigh importance

Anthropic releases FlashProve: an open-source “provable inference” prototype with zkSNARKs

Anthropic’s Frontier Red Team published progress on cryptographic proofs (zkSNARKs) to verify that a response came from a committed model without revealing the weights, and released FlashProve under Apache 2.0 (code, C++ and Lean verifiers, Lean proofs, and docs). On Llama 4 Maverick (400B MoE, 4-bit quantization), a single H100 generates proofs at about 110 tokens per second, with about 19 MB per 1,024-token request — about 70 to 95 times the GPU time of vLLM on the same model; verification takes about 17 seconds on 24 CPU threads. The post lists limits: no formal guarantee on the fast executable code, proof over an integer model under greedy decoding, architecture not hidden, contexts up to 1,024 tokens. @AnthropicAI amplified the announcement on X on Tuesday.