Anthropic Open-Sources FlashProve for Zero-Knowledge LLM Inference Verification
Model substitution by API providers has been a trust problem with no cryptographic answer until now; FlashProve turns 'which model am I actually calling' into a checkable claim. Proof generation is still 72–94x inference cost, so it suits spot audits rather than per-request verification today. Teams building on third-party APIs — especially those serving regulated markets in Singapore and wider ASEAN — can use the open-source verifier to hold providers accountable.
Anthropic released FlashProve, an open-source system that uses zero-knowledge proofs to verify that API responses come from a committed model rather than a silently substituted cheaper one. The verifier needs neither model weights nor GPUs; an ordinary CPU can check a proof in roughly ten-plus seconds, with each 1024-token request producing about a 19MB proof. Anthropic tested it on Meta's Llama 4 Maverick, a roughly 400B-parameter MoE with about 17B active parameters per token, generating proofs at around 110 tokens per second on a single H100. The main cost is high: proof generation consumes about 72 to 94 times the GPU time of normal inference. The prover, verifier and formal proof code are all open-sourced.
当你调用某个大模型API时，怎么证明真正回答你的就是商家说的那个模型？怎么知道有没有被偷偷换成别的便宜模型？
Anthropic刚开源了FlashProve，一个基于零知识证明的大模型推理验证系统，让服务商为模型推理结果生成密码学证明，验证输出确实来自事先承诺的模型权重
以后AI服务商给你说用的哪个模型，你可以验证
而且，验证者不需要拿到模型权重，也不需要GPU，只用普通CPU就可以检查证明
Anthropic用Meta的Llama 4 Maverick做了实验，总参数约400B、每Token激活约17B的一个MoE
单张H100大约以110 token/秒生成证明；每1024 token请求的证明约19MB，普通CPU大约十几秒可以验证
成本仍然是高的，生成证明大约消耗正常推理72～94倍的 GPU时间
FlashProve的证明生成器、验证器和相关形式化证明代码全部已开源
#FlashProve #可验证模型 #可信AI
