Anthropic released FlashProve, an open-source system for provable inference. It uses zkSNARKs to let users verify a response came from a specific model. A single H100 GPU generates proofs at 110 tokens per second for a 400B-parameter model. Proofs are about 19 MB per 1,024-token request. The system is 70-95 times slower than running inference on vLLM. It hides model weights but not architecture. Lean proofs audit the security guarantees. The release includes code, a technical report, and a protocol description. Applications include pacing AI development and detecting tampered responses. Regulators could use it to enforce approved-model rules.
What changed
Users could not verify which model produced a response or detect tampering in the token stream.
What it unlocks
Verifying that a response came from a specific committed model without re-running inference or exposing weights.
- 110 tokens/second proof generation
- ~19 MB proof size per 1,024 tokens
- 70-95x GPU time vs vLLM
- 400B-parameter Llama 4 Maverick model
What you need to act on it
- a zkSNARK-compatible model setup
- access to FlashProve code and proofs
Sources