Recursive Flint: A Laptop Coding Model Built Largely by AI

Recursive, the AI lab Richard Socher started in 2025, has released its first model, and the pitch is as much about who built it as what it does. Flint is a coding model for agents that runs on a laptop, and Recursive says its automated research system, the “Eureka Machine”, helped with nearly every step of making it, from the architecture and the training recipe to the inference kernels. A small team of people and agents built it from scratch in under four months, according to the company. The weights are on Hugging Face under the Apache 2.0 licence.

Flint is a mixture-of-experts model with 39.8 billion parameters in total, of which about 2.5 billion are active for each token: 256 routed experts per layer, six chosen per token, plus a shared one. The model card documents a 262,144-token context window. The shipping version is quantised to NVIDIA’s 4-bit NVFP4 format, which Recursive says keeps it within 0.7 points of the 16-bit model across its benchmarks. On an M4 Pro MacBook Pro it generates more than 125 tokens a second, and more than 150 on a 64GB RTX Spark, measured over a 128,000-token session. TechCrunch puts that at roughly twice the 58 to 69 tokens a second Recursive measured for Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning. It runs in llama.cpp, Ollama, LM Studio, vLLM and SGLang, and Recursive has published its own Mac inference engine, rec.

What the Eureka Machine did

Recursive is specific about the machine’s contributions. It searched expert counts and widths starting from a 35-billion-parameter, 3-billion-active baseline and settled on the 40B/2.5B shape as the best trade between speed and coding quality. It found a quantisation recipe that pushes even the output layer down to 4 bits, against common practice, and then added a distillation stage, with the 16-bit model as teacher, that roughly halved the training steps compared with standard quantisation-aware training. And it wrote the low-level code: kernels for 4-bit projections, expert routing and attention, plus a Metal “megakernel” that fuses Flint’s whole decoding pass on Apple’s M4 and M5 chips. What the company has not said is how much researcher time any of that saved, a gap TechCrunch also noted. Socher’s stated goal, described in an earlier Recursive post and in his launch video, is a system that improves AI on its own under human supervision, which makes Flint an early, deliberately narrow test of the idea.

Where it lands on the benchmarks

Recursive’s claim is that Flint leads mixture-of-experts models under 50 billion total parameters on agentic coding, and its own chart supports that: it scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, ahead of Alibaba’s Qwen 3.5 and 3.6 and Laguna XS 2.1 on all three, though by only 0.8 points over Qwen 3.6 on SWE-bench Verified. Against dense models of 27 to 31 billion parameters the picture is mixed. Qwen 3.8 beats it on all three agentic tests, by 18 points on Terminal-Bench, and Meta’s Muse Glimmer edges it on SWE-bench Verified. Flint’s argument against those models is speed: with about a sixteenth of its weights active per token, it answers several times faster than a dense model of that class.

Recursive is candid that Flint needs more direction than the paid cloud assistants, and suggests breaking large tasks into smaller steps. Wired’s Will Knight, who tried it on a 48GB Mac to build a to-do app and a web research tool, found that it handled every request, according to TechCrunch, where other small coding models he tested stalled or erred. The release lands a week after Meta’s Muse Spark Open, an 816-billion-parameter model that needs about a terabyte of memory; Flint makes the opposite bet, small enough to keep code on the developer’s own machine with no usage bill. For Recursive, which came out of stealth in May with $650 million in funding, the bigger claim is the method. Setup instructions are in its documentation; whether the Eureka Machine can do the same at frontier scale is the question Flint leaves open.