HeyHomees Lab Notes · No. 16 · Anthropic · Published November 4, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Proves model output without exposing weights
- Runs full-size Llama 4 Maverick on one NVIDIA H100, ~110 tokens/sec prefill
- Verification can be done on a regular CPU in seconds with cp_verify
What is it?
FlashProve is a research artifact that enables an AI provider to prove which model generated a response while keeping the model’s weights secret. It is released as version 0.16.0 and is licensed under Apache 2.0.
How does it work?
The provider first publishes a commitment ID that binds to a specific weight set. For each reply the GPU creates a cryptographic proof that the output matches the committed model, and anyone can verify the proof on a CPU using cp_verify in seconds. The protocol’s soundness is formally proved in Lean 4, with a cheating probability of about q/2^124 + 2⁻¹⁰¹ after q hash queries.
What does it need to run?
It runs the full-size Llama 4 Maverick with a 1,024-token context on a single NVIDIA H100 GPU. Verification requires Linux, bash, g++ 13+ and under 2 GB RAM; building the prover also needs typical system packages.
By the numbers
- Version: 0.16.0
- Model: Llama 4 Maverick (full size)
- Context window: 1,024-token
- Hardware: single NVIDIA H100
- Prefill speed: about 110 tokens/second
The honest read
The listed speed of about 110 tokens/second is reported by the FlashProve authors and reflects pre-fill performance on a single NVIDIA H100 at a batch size of 24; no external benchmarks are provided. The soundness guarantee (probability about q/2^124 + 2⁻¹⁰¹) is a theoretical bound proved by the authors in Lean 4, not an empirical measurement of attack resistance.
What we don't know yet
- Exact download size of the repository
- Real-world CPU verification latency beyond “seconds”
- Support for models other than Llama 4 Maverick
Sources
- README, GitHub (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.