Recursive Flint runs autonomous research on a laptop, open-sources 40B-A2B

Flint by Recursive is an automated research system that runs on a local laptop. Useful automated research has typically been tied to the largest closed-source proprietary models, but congml found in hands-on testing that Flint can already complete a full loop — proposing ideas, writing code, running experiments, and improving results — on a single laptop, with every experiment inspectable and the user choosing what it researches next.

Confirmed

• Background: The author opens with OpenAI's Navier–Stokes solver and AlphaEvolve finding faster algorithms, arguing that AI doing research is becoming very real and that Flint makes this capability locally accessible; Josh Tobin was reportedly surprised (via a repost) at how well autoresearch works on small models.
• Task 1 (bike demand forecasting): Using the public UCI Bike Sharing dataset, Flint predicted rental demand from weather and calendar data, automatically combining improvements such as "log transform + quadratic time trend + boosted residual trees," cutting validation RMSLE by 30.8% after 3 model-hours.
• Task 2 (job-shop scheduling): Flint wrote its own methods for reordering operations and searching for better schedules; with the same jobs and machines, completion time beat the round-robin baseline by 26%. The whole pipeline required no visual input — only code and score iteration.
• Task 3 (circle packing): Placing circles in a square to maximize the sum of radii, Flint pushed the total radius of 26 circles to 2.6294 after 128 experiments using strategies like code that moves circles in coordination.

Why it matters

• The three tasks span prediction, optimization, and geometric search, showing that autonomous research capability is moving from the largest closed models down to locally runnable small models, with inspectable experiments and room for human intervention — lowering the barrier to reproduction and verification.

2026-10-22 ~ 2026-10-22 · 7 related posts

• Running an autonomous research loop locally with Recursive Flint, from scheduling to bike-demand forecasting — cong_ml · 2026-10-22
• Local auto-research with Flint: 128 experiments push circle packing to 2.6294 — cong_ml · 2026-10-22
• [source] Flint runs auto-research on a laptop: 26% faster scheduling vs round-robin — cong_ml · 2026-10-22
• [source] Flint cuts bike-demand validation RMSLE 30.8% in 3 hours of autonomous model iteration — cong_ml · 2026-10-22
• Local autoresearch on a laptop: Flint proposes ideas, writes code, tests and improves results — cong_ml · 2026-10-22
• Autoresearch runs surprisingly well on tiny local model Recursive Flint — cong_ml · 2026-10-22
• [source] Recursive Flint-40B-A2B runs an AI scientific discovery loop locally on a laptop — cong_ml · 2026-10-22