Recursive debuts Flint, a laptop coding model it says AI largely built
What's the deal? Recursive on Wednesday released Flint, a free coding model that runs on a laptop and was built largely with the help of AI. The startup says its automated research system, the Eureka Machine, contributed to nearly every stage of the model's development, letting a small team build it from scratch in under four months.
Why it matters: Flint is an early public test of recursive self-improvement — the contested idea that AI can help build better AI under human oversight. "Nearly every aspect of the entire model pipeline was actually done and improved upon by AI," said Richard Socher, Recursive's chief executive officer and co-founder.
The specs: Flint is a mixture-of-experts model that activates just 2.5 billion of its roughly 40 billion parameters per token, making it light enough for an ordinary laptop. It is free to download on Hugging Face under the Apache 2.0 open source license.
In Recursive's tests, Flint generated about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip — roughly twice the speed of comparable models — and topped 150 tokens per second on Nvidia's RTX Spark. On SWE-Bench Verified, a benchmark for fixing real-world bugs, it outperformed other models under 50 billion parameters.
Under the hood: On its Hugging Face model card Recursive reports scores of 74.2% on SWE-bench Verified, 54.9% on Terminal-Bench 2.1 and 89.7% on SWE-bench Pro v2, and says Flint supports context windows of up to 512,000 tokens, with 256,000 recommended. The company says its Eureka Machine searched for the model's sparse shape (40 billion total, 2.5 billion active parameters, chosen over a 35B/3B baseline), then went against common practice by quantizing the output layer to 4 bits and recovering accuracy through self-distillation, leaving the final NVFP4 model within 0.7 points of the 16-bit version. It also wrote custom kernels for Apple Silicon and Nvidia hardware, including a Metal megakernel that fuses the whole decoding pass on M4 and M5 chips. Recursive is shipping its own lightweight inference engine, rec, alongside support for llama.cpp, Ollama, LM Studio, vLLM and SGLang.
Why local? Because Flint runs on a user's own machine, developers can work offline, keep their code private, and avoid the usage fees cloud AI services charge. Running it on Apple Silicon requires an M4 Pro, M5 Pro, or M5 Max chip and at least 48GB of memory.
What could go wrong? Recursive acknowledged that Flint needs more guidance than large paid coding assistants, advising users to break big tasks into smaller steps. Its practical standing will depend on independent evaluations and real-world use, not company benchmarks alone.
The backstory: Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, backed by GV, Nvidia, and AMD VenturesDealroom has a profile for this one. Try Dealroom →. It has also secured a $410 million compute deal with Amazon Web ServicesDealroom has a profile for this one. Try Dealroom →, which it framed as focused on deploying more AI agents rather than hiring more researchers.
The signal: Flint enters a crowded local-model market, with recent coding releases from Qwen, Meta, and Nvidia. Its real distinction is less that it runs locally than Recursive's claim that automated research built it in under four months — an early, still company-reported, test of AI-assisted model development.
Read more: TechCrunch · Recursive: Introducing Flint · Hugging Face model card
Image credit: Generated with Gemini
