Back to events
Safety
Score 82.3Heat 0

Anthropic releases FlashProve for provable inference in large language models

ResearchModel SecurityOpen WeightsEvaluationLlama 4 MaverickAnthropicFlashProve

Summary

Anthropic has released FlashProve, an open-source cryptographic proof system utilizing zkSNARKs for provable inference on large mixture-of-experts models. The system generates proofs at approximately 110 tokens per second on a single H100 GPU while keeping model weights hidden.

Why it matters

Provable inference provides cryptographic verification that a specific AI model generated a response, improving trust, tamper detection, and regulatory compliance.

Editorial score

2 reviews
8
Significance
8.5
Novelty
9
Credibility
7
Reach
7.5
Actionability