FlashProve
Anthropic released FlashProve, an open-source research prototype that uses zkSNARK proofs to show a model response came from committed hidden weights, with proving at 110 tokens per second on Llama 4 Maverick.
Providers publish a short commitment to an exact set of weights, and each prompt and response pair gets a proof that anyone can check on a CPU without the weights. The prover uses Maverick's mixture-of-experts structure, so work scales with about 17 billion active parameters and expert choices stay hidden. It is a software-only design that does not rely on trusted hardware.
- Date
- Monday 2 November 2026
- Lab
- Anthropic
- Kind
- infra
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Proof generation speed, Llama 4 Maverick | 110 tokens per second reported by Anthropic | company |
| Batch of 24 requests proved on one H100 with 24 CPU cores | about 223 seconds 4-bit conversion of the model | company |
| Proof size per request | 18.78 MB | company |
| Verification time | 17 s with 24 threads, 21 s with one thread CPU only, about 6.5 MB commitment | company |
| Proving cost versus FP8 vLLM on eight H100s | about 72x per-GPU time, about 94x at highest vLLM load tested | company |
Code is under Apache 2.0 but labeled an unmaintained research artifact (v0.16.0), with no independent audit, 1,024 token context limit, and greedy decoding only.
Sources
This record was partly confirmed: some claims could not be checked on 3 November 2026. How we check