AI Card Bao BCAST.AICARDBAO.COM

Industry News

Spending all that money just to use a scaled-down version of the model? Anthropic’s open-source tool FlashProve can determine whether this is the case in just 17 seconds.

Many people who use third-party APIs have wondered why the answers provided by these models, which are supposedly based on flagship versions, seem to become increasingly weak or less accurate over time. However, there has never been any concrete evidence to support this claim, because it is impossible to obtain the underlying weighting parameters used by these algorithms or to run the experiments again under the same conditions.

This week, the Anthropic research team released FlashProve under an open-source license, choosing to pursue the path of zero-knowledge proof technology.

The process consists of three steps. Before the model is deployed, approximately 39 minutes are required to generate a public “fingerprint” for the weights on a single H100 device. After each inference session, the GPU automatically generates a proof file of around 19MB. Users can then use this proof file to perform verification on an ordinary CPU; the verification takes only 17 seconds to complete. Throughout this entire process, neither the weights nor a graphics card are required.

The challenge lies in the scale. The model used for testing was Llama 4 Maverick, which contains 400 billion parameters; even after 4-bit quantization, it still occupies 232GB of memory—while a single H100 graphics card only has 80GB of video memory. They took advantage of the design of hybrid expert models: in practice, only around 17 billion parameters are activated for each token, meaning that the amount of data required for training is directly related to this subset of parameters. The weights for these parameters are fed to the GPU in segments via a streaming mechanism, which helps to optimize memory usage.

What’s even more interesting is how this code was actually written. Approximately 95.2% of the code stored in the repository consists of Lean 4 formalized proofs, with a total size of around 398MB. The 114-page document detailing the implementation, as well as the underlying code itself, were mostly generated by Claude. Humans only participated in finalizing six security-related definitions; the error rate during this process was lower than one part in a trillion, and the likelihood of any fraudulent alterations was even smaller than one part in 2 to the power of 100. The rest of the validation process was entrusted to theorem provers.

Don’t get too excited just yet. FlashProve is currently just a research outcome and there are no plans for its commercial deployment. The fact that each verification request requires 19MB of data means that frequent use would place a huge burden on both transmission and storage services provided by intermediaries. Moreover, it’s highly unlikely that the target parties would be willing to cooperate in this regard. For FlashProve to be put into practical use, some major client would need to include the requirement for “verification documents” in their purchase contracts.

The scope of what it can guarantee also has its limits: what it can prove is merely that this particular output is indeed derived from those specific weights. There is no detailed information in the public materials regarding whether inputs such as system prompts have been manipulated in any way.

Related