Recursive Releases Flint, an Apache-2.0 Open-Weight Coding Model for Laptops That It Says Its Automated-Research System Helped Build
Recursive Superintelligence released Flint, a 40B-total, 2.5B-active mixture-of-experts coding model, with weights downloadable from Hugging Face under Apache 2.0. Recursive says Flint leads MoE models under 50B parameters on three agentic coding benchmarks and runs at more than 125 tokens per second on an M4 Pro MacBook Pro, and that its Eureka Machine contributed to nearly every part of the build. All benchmarks are Recursive's own.
The first model from a lab trying to build AI that improves itself is a coding model small enough for a laptop, and its weights are openly licensed and downloadable. On Wednesday, October 21, Recursive Superintelligence released Flint, a mixture-of-experts model with 40 billion total parameters that activates 2.5 billion per token. Recursive says its "small team of people and agents built it from scratch in under four months."
What we verified about the weights
Recursive says Flint's weights are on Hugging Face under the Apache 2.0 licence. When we checked at 02:58 UTC on October 22, the RecursiveSI/flint repository was public and not gated, listed an Apache 2.0 licence and included the licence text, and held 15 safetensors shards totalling about 39.8 billion parameters by Hugging Face's count; a download of one shard began normally. The model card says hosted API terms are separate from the weight licence. Recursive also published int4, NVFP4 GGUF and int4 GGUF variants.
Recursive's claims
Recursive says Flint leads mixture-of-experts models with fewer than 50 billion total parameters on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified, and generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB Nvidia RTX Spark "at short contexts." It supports contexts up to 512,000 tokens, with 256,000 recommended. Its agentic scores are averages of pass@1 over at least five attempts on its own harness settings. TechCrunch reports that on Recursive's six-test average, only Alibaba's Qwen3.8 27B scored higher than Flint.
Recursive says its Eureka Machine, a system that proposes, implements and tests research ideas, contributed to Flint's data, training recipes, infrastructure and inference code. Among its examples: an architecture search that settled on the 40B/2.5B shape, a 4-bit NVFP4 quantisation recipe it says scores within 0.7 points of the 16-bit model, and custom kernels for Apple Silicon and RTX Spark. TechCrunch describes Recursive's approach for now as "an AI research assistant overseen by people," and reports the company emerged in May with $650 million at a $4.65 billion valuation; its CEO is Richard Socher.
Limits
- Every benchmark and speed figure is Recursive's own, on its own harnesses and hardware.
- Independent check: Artificial Analysis's model leaderboard did not list Flint when we checked at 02:59 UTC on October 22.
- TechCrunch notes that Recursive did not say how much time the Eureka Machine saved its researchers.
Sources
- Recursive: Introducing Flint (October 21, 2026)
- Hugging Face: RecursiveSI/flint model repository (accessed October 22, 2026)
- TechCrunch: Recursive launches Flint, a laptop coding model it says was largely built by AI (October 21, 2026)
- Artificial Analysis: model leaderboard (accessed October 22, 2026)