Recursive Releases Flint, a Free Coding Model That Runs on a Laptop and Was Largely Built by AI

On quality the picture is more modest. By Recursive's own tables, Flint scores 74.2 on SWE-bench Verified, behind Alibaba's dense Qwen 3.8 at 75.6 and Meta's dense Muse Glimmer at 76.0. It leads the small mixture-of-experts models, but i-scoop notes its Terminal-Bench result is its weakest and that Recursive's SWE-bench Pro v2 score is its own measurement, not comparable with the public leaderboard.

What did the AI actually build?

Recursive says its automated research system, the Eureka Machine, helped with nearly every step, from pretraining to inference, and that a small team shipped Flint in under four months. Credited findings include a sparser model shape, low-bit training that shrank the model by 71% with almost no quality loss, and custom kernels, TechCrunch reports. i-scoop adds that its four-bit version scores within 0.7 points of the full-precision one.

Humans supervise the system and can stop it. Recursive has not said how much researcher time or money it saved, so the claim that AI built most of Flint is a statement of method, not a measured saving.

Should developers switch?

Wired's Will Knight found Flint reliable on coding requests, while smaller models often produced errors or ran slowly, TechCrunch reports. Recursive itself advises splitting large tasks into small steps, since Flint needs more guidance than big paid assistants. For private code on a well-specified job, that trade may be worth it.

The next signal will come from independent tests. If outside benchmarks match the company's tables, a laptop-class model built with so little human effort will raise the pressure on cloud-only coding tools.