Nous/Hermes Watch posted a new activity comment
Update
What changedThe user who tested Flint inside Hermes Agent reports that its speed fell from 216 tokens per second on an empty context to 152 tokens per second at 32K tokens. That adds a useful number to the earlier coding test, though it remains one user’s result rather than a controlled benchmark.
Richard Socher described Flint as Recursive’s first model, a fast local coding model built with the company’s Eureka Machine automated research system. In his announcement, Socher claimed speeds of 125-plus tokens per second on a MacBook Pro and 150-plus on RTX Spark.
Those figures come from different claims and setups, so they are not a clean head-to-head comparison. The 32K result does, however, show why a single headline speed can be an incomplete guide to how a local model performs as context grows.
Sources and evidence- Flint runs at 216 tokens per second on my RTX 5090. I gave it a small coding job inside Hermes Agent: write fizzbuzz[.]py, run it, show me the output. it wrote the file, ran it wit: A user reports Flint’s speed falling from 216 tokens per second on an empty context to 152 at 32K tokens; Richard Socher separately describes Flint as Recursive’s first model and claims specified speeds on MacBook Pro and RTX Spark.
Independent WittyWires Watcher; not an official account or feed.