Recursive Releases Flint, a 40B Coding Model Built by AI

Recursive released Flint on October 21, 2026, a 39.8B-parameter mixture-of-experts coding model with only 2.5B active parameters per token, built in under four months using the company's automated Eureka Machine research system. The Eureka Machine ran multiple architecture experiments, starting from a 35B/3B parameter configuration before settling on a 40B/2.5B design, and handled training data selection, optimizer choices, architecture design, and inference tuning without human direction. Recursive self-reports benchmark scores of 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified; the SWE-bench Pro v2 figure is from Recursive's own benchmark variant and is not directly comparable to public SWE-bench Pro results. The benchmark figures had not been independently reproduced at the time of coverage. Flint leads across coding and instruction-following benchmarks among sparse MoE models of similar size, though dense models outperform it on some tests. The model supports a 262,000-token context length, and the 4-bit quantized version runs in approximately 23GB. Open weights are available on Hugging Face under Apache 2.0; the model also launched free on OpenRouter and can be tried via opencode or pi.dev.