Roy Rinberg open-sources FlashProve, a prototype proving 400B model inference

Roy Rinberg has open-sourced FlashProve, a research prototype that proves a 400B-parameter model runs at over 100 tokens per second on one H100 in prefill.

The announcement, posted on X, says the release is meant to encourage further progress in the field.

The report attributes the release to Anthropic, while the subject of the fact is Rinberg. The reports do not give further detail on the prototype's method or independent verification of the throughput figure.