Recursive, a lab trying to automate AI research itself, has released Flint, a coding model built to run on your own computer instead of in a data center. Flint holds 40 billion parameters in total, but only 2.5 billion of them work on any given piece of text. The weights are free to download, according to Recursive’s launch post.
That gap between total and active is the design. Flint is a “mixture-of-experts” model, which splits its knowledge across many specialist sub-networks and wakes only a few of them for each chunk of text. A hospital works the same way: hundreds of doctors on staff, but your visit involves two or three. The model carries the knowledge of a large system and pays the running cost of a small one.
Low running cost is what makes laptops viable. On a MacBook Pro with an M4 Pro chip, Recursive says Flint produces over 125 tokens each second (a token is a word or a piece of one). With NVIDIA’s RTX Spark (the 64GB version), it quotes over 150, with short contexts in both cases. The company claims a lead of up to 2x against comparable models. Against “dense” models in the 25 to 40 billion range, where every parameter works on every token, it claims a gap of more than 5x at similar coding scores.
Those scores are the company’s own. Recursive says Flint tops all other mixture-of-experts models that stay under 50 billion parameters overall, judged on coding-agent tests. It names SWE-Bench Verified, SWE-Bench Pro v2, and Terminal-Bench 2.1. The launch post includes no independent results. It is unusually frank about the limits, though: Recursive says Flint may need extra hand-holding compared with the big closed coding models developers know, and advises splitting complex jobs into smaller subtasks with explicit instructions.
The build story is the other claim. Recursive says its team was small, made up of people and agents, and finished Flint in under four months without starting from an existing model. The agent side is the Eureka Machine, a system that proposes research ideas, implements them, tests them, and uses the results to pick the next experiment. Per the post, it began with a design of 35 billion total and 3 billion active parameters, tried different expert layouts, and settled on the 40 billion and 2.5 billion split. It also searched for ways to shrink the model after training. The compressed version lands 0.7 points short of the full-precision one on Recursive’s own tests, the company says.
Access is wide. Flint is on Hugging Face with an Apache 2.0 license, a permissive license that allows commercial use and modification. Recursive lists support for Ollama, LM Studio, llama.cpp, Unsloth, SGLang, and vLLM, and it released a lightweight engine called rec, tuned for recent Apple chips. Context windows reach 512,000 tokens, though the company suggests staying under 256,000.
The pitch to developers is practical: work offline, keep code on your own machine, pay no API fees. That appeals most to teams whose code cannot leave the building.
The speed figures carry a qualifier. The 150 number applies at short contexts, and agentic coding sessions tend to grow long. The post’s chart note mentions an average across a 128,000-token session, but the text states no figure for it, so outside testers will first check how fast Flint stays once the conversation gets big.
Reported by Recursive in its launch post “Introducing Flint” (undated on the page).
