Socher launches Flint, an open-weights local coding model

Richard Socher introduced Flint, the first model from @Recursive_SI, built with what he calls the Eureka Machine, an automated AI research system. It is a local coding model, which means it runs on your own hardware rather than against an API meter. Open weights are up on Hugging Face now, and he says you can also try it free on opencode or pi.dev.

Richard Socher

@RichardSocher

X

Private, unmetered intelligence on your own machine.

Richard Socher

@RichardSocher

X

This is what allowed our very small team to build this model in <4 months.

The speed numbers are the headline. Socher says Flint hits 125 or more tokens a second on an M4 Pro MacBook Pro and 150 or more on a 64GB NVIDIA RTX Spark. He claims that is up to twice as fast as comparable models on short contexts, and more than five times faster than dense models with similar coding scores. On quality, he says Flint is the best performer in its class, which he defines as mixture-of-experts models under 50B parameters, on agentic coding benchmarks including SWE-bench Pro v2, Terminal-Bench 2.1 and SWE-bench Verified.

Four months, small team

The part worth sitting with is the build story. Socher says the whole thing went from scratch to shipped in under four months, and credits the Eureka Machine for making that possible with a very small team. He describes it running the full research loop, ideation, implementation and validation, then using the results to pick what to try next. Among the things it contributed, he lists a sparser model shape, low-bit training that shrank Flint by 71% with almost no loss in quality, and custom kernels.

All of these are the vendor's own figures. Nobody independent has published a verdict yet, and "best in class" is doing a lot of work in a category the vendor also defined. Treat the benchmark table the way you would treat any launch day chart.

The business read

If the speed claims hold up on your machine, the interesting change is not the leaderboard, it is the cost floor. Plenty of solo founders now run agents through the day on metered API calls for refactors, test writing, log reading and other grunt work where the token bill is real but the output is not precious. A local model that is fast enough to keep up moves that tier of work to roughly zero marginal cost, and keeps the code on your laptop. Socher leans on exactly that framing.

It runs offline through ollama, LM Studio, Unsloth, exo, vLLM, SGLang and llama.cpp, so trying it is a download rather than a signup. The thing to check is whether 125 tokens a second on an M4 Pro is still 125 when your context is long and your repo is messy.