Richard Socher's Lab Let An AI System Design Its First Model

Flint carries 39.8 billion total parameters and activates roughly 2.5 billion per token. That ratio is the whole design: about 6% of the model does work on any given step, which is how a 39.8B model ends up running at laptop speed. Richard Socher puts it above 125 tokens per second on a MacBook Pro and 150 on NVIDIA's RTX Spark, and the checkpoint is released for local coding-agent use.

The throughput number is the one that matters, because for a coding agent speed is not a comfort feature. An agent that reads files, proposes an edit, runs a test and reads the failure generates many thousands of tokens per task — at 20 tokens per second that loop is unusable, at 125 it is a working tool. Running on a laptop also means no API bill, no rate limit and no code leaving the machine, which is the actual reason most teams wanting local coding agents want them.

The claim that will get argued about is that Flint was built by an automated AI research system rather than a human research team, which Recursive presents as the point of the company. Be precise about what that does and does not assert: it says an automated system did the architecture and training work producing this checkpoint, not that it invented the underlying techniques, and sparse mixture-of-experts at this ratio is established practice. The interesting question is whether the loop produces a better model next time without more human input — and that gets answered by Recursive's second model, not this one. No benchmarks, license terms or comparisons accompanied the release, so Flint is currently a throughput claim with a checkpoint attached: worth downloading and testing, not yet worth switching to.

Sources

The Neuron daily digest