Anthropic open-sources FlashProve provable inference prototype
Original title: Anthropic open-sources FlashProve, its provable inference prototype for MoE models
FlashProve cryptographic proofs confirm specific models generated specific replies, tested on Llama 4 Maverick MoE model.
Anthropic Frontier Red Team with EPFL, MIT and StarkWare collaborators released FlashProve prototype producing cryptographic proofs that one LLM generated one reply. Repository includes prover, C++ verifier and Lean reference verifier under Apache 2.0 license. Tested on Llama 4 Maverick 400 billion parameter Mixture-of-Experts model requiring one 80 GB GPU.
- Prover, C++ verifier and Lean verifier released
- Proof for 1024-token request ~18.78 MB
- Checking takes ~17 seconds on 24 CPU threads
- Offline commitment ~39 minutes per model
- 72x overhead vs vLLM
- Vendors
- Anthropic
- Products
- FlashProve
- Organizations
- EPFLMITStarkWareBroad Institute
- AI models
- Llama 4 Maverick
From the original article
2 min read at helpnetsecurity.com
Anthropic’s Frontier Red Team, with collaborators at EPFL, MIT and StarkWare, built FlashProve, a prototype that produces a cryptographic proof that one specific language model generated one specific reply. The prover, two verifiers, the formal proofs behind them and a protocol specification are on GitHub under the Apache 2. 0 license.
By Anamarija Pogorelec · helpnetsecurity.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.