Recursive releases Flint, a free open-weight 40B-parameter mixture-of-experts coding model that runs on laptops, built with its AI research system

Recursive published Flint, its first model, on Hugging Face under Apache 2.0. It is a mixture-of-experts coding model with 40B total parameters, of which 2.5B are active per token. Recursive reports about 131 tokens/s on a MacBook Pro M4 Pro, against 58-69 tokens/s for similar models such as Gemma 4 and Nemotron 3.5 Lightning, and over 150 tokens/s on Nvidia RTX Spark. It also says Flint beats other sub-50B models on SWE-Bench Verified. On its own six-test average only Qwen3.8 27B scored higher. The speed and benchmark figures are Recursive's own, and only Wired's hands-on use backs them up independently.

Developers can download the model and test it offline, and the permissive licence allows commercial use. That matters most to teams who want a private coding assistant, or who want to cut cloud assistant costs. The weights are public, so the speed and benchmark claims can be checked by anyone. Whether it is good enough to replace paid tools for real work is not yet shown. Competitors in small open models now face a direct speed comparison.

The "early public test of recursive self-improvement" framing is company packaging and cannot be verified from outside. By Recursive's own account, humans oversaw an AI research assistant, which is closer to automated tooling than AI building better AI unsupervised. The "under four months with a small team" claim has no baseline to compare against. The benchmarks are also self-reported, and the model loses on Recursive's own six-test average. No primary evidence links were captured in the extraction.

1. 01Independent reproductions of the 131 tokens/s speed and the SWE-Bench Verified result, from third-party evaluators or community runs on comparable laptops, within the next few weeks.
2. 02Hugging Face download counts and real-world developer reports on code quality, long-context behaviour and failure rates compared with Qwen3.8 27B and Gemma 4.
3. 03Whether Recursive publishes a technical report detailing what Eureka Machine did at each stage, and whether others can replicate the 71% size reduction from low-bit training.