OPEN LINK ↗
// 1h agoOPENSOURCE RELEASE
Anthropic FlashProve Proves Model Provenance
Anthropic released FlashProve, an Apache 2.0 research prototype using zkSNARKs to verify that responses came from committed model weights without revealing them. On Llama 4 Maverick, it generates proofs at roughly 110 tokens per second, with 19 MB proofs verifiable in seconds.
// ANALYSIS
FlashProve makes model provenance technically tangible, but its high proving overhead and narrow prototype constraints keep it out of production for now.
- –Targets model swapping, tampered inference, and third-party router accountability
- –Demonstrates a major speedup over earlier zkML systems on a 400B-parameter MoE model
- –Proof generation still costs roughly 70–95 times standard inference
- –Current verification covers only 1,024-token contexts, integer weights, and greedy decoding
- –Open-source code and Lean proofs could accelerate independent research and auditing
// TAGS
flashprovellminferencesecurityresearchopen-source
DISCOVERED
1h ago
2026-11-03
PUBLISHED
1h ago
2026-11-03
RELEVANCE
9/ 10
AUTHOR
AnthropicAI