Recursive, the AI startup trying to build systems that improve themselves, released its first model on Wednesday. Flint is a free, open-weight coding model small enough to run on a laptop, and the company says it was built largely by AI: its automated research system, the Eureka Machine, helped with nearly every step of Flint’s development, letting a small team build the model from scratch in under four months, TechCrunch reports.
Flint is downloadable from Hugging Face under the Apache 2.0 licence. Because it runs on the developer’s own machine rather than a cloud service, it works offline, keeps code private and carries no per-token usage fees.
Flint: A 40B Mixture-Of-Experts Built For Agentic Coding
Flint is a mixture-of-experts model with about 40 billion total parameters and roughly 2.5 billion active per token, with a context window of up to 512K tokens, 256K recommended. Recursive says it generates more than 125 tokens per second on an M4 Pro MacBook Pro, and more than 150 tokens per second at short contexts on an NVIDIA RTX Spark system. The full 16-bit checkpoint is about 80 GB and runs on a single data-centre GPU; smaller quantised versions are published for Nvidia and Apple Silicon hardware, and the model works with llama.cpp, Ollama, LM Studio, Unsloth, vLLM and SGLang, plus Recursive’s own lightweight engine, rec, tuned for Macs with M4 Pro and M5 Pro or Max chips.
On benchmarks, Recursive reports 74.2 percent on SWE-bench Verified and 54.9 percent on Terminal-Bench 2.1, and says Flint leads mixture-of-experts models under 50 billion parameters on agentic coding benchmarks and matches dense models of 25 to 40 billion parameters while running more than five times faster. Those are company-reported numbers, run under Recursive’s own setups, and independent testing will decide how much weight they carry. Early hands-on impressions reported by AI Parabellum were positive, with the caveat that Flint may need clearer instructions than the large closed models for complex jobs.
Flint: What The Eureka Machine Actually Did
The bigger claim is the production process. Recursive says the Eureka Machine proposes ideas, writes the code to test them, checks the results and decides what to try next, running many experiments in parallel under human oversight. For Flint it picked the architecture, settling on the 40B-total, 2.5B-active expert layout; it found a quantisation approach that shrinks the output layer to 4-bit precision while staying within 0.7 points of the full model on benchmarks; it wrote custom kernels tuned for Apple Silicon and NVIDIA RTX Spark chips; and it worked on training data, optimisers and recipes.
“Nearly every aspect of the entire model pipeline was actually done and improved upon by AI,” chief executive Richard Socher told Wired, saying that is the only reason a handful of researchers could compete. Socher also says the system that built Flint cannot escape human control, and that any future fully self-improving version would still be overseen by people who can stop it.
Why It Matters
Flint matters less as a coding model, a crowded field that already includes tools surveyed in this site’s guide to AI coding tools, than as the first public test of recursive self-improvement as a production method. If an automated research system genuinely compresses a frontier-adjacent model build into under four months with a small team, the economics of model development change for every lab that does not own a hyperscale data centre. The honest caveat is built into the story: Recursive is both the vendor of the Eureka Machine and the judge of how much credit it deserves, its benchmark and speed figures are self-reported, and the real test is what developers find when Flint runs on their own laptops next to models built the conventional way, including local-first efforts like NVIDIA’s Nemotron open models for always-on agents. Flint’s release at least makes that test possible today, in the open, which is more than most self-improvement claims offer. The claim also lands in the same week as Google’s Harness2 framework for agents that rewrite their own harness.
Background
Recursive was founded last year and emerged from stealth in May 2026 with $650 million in funding at a $4.65 billion valuation, backed by GV, Nvidia and AMD Ventures. It is led by Richard Socher, Salesforce’s former chief scientist and a well-known natural-language-processing researcher. Flint is the company’s first product and the first public artefact of its self-improvement thesis.
Source: TechCrunch, AI Parabellum
0 Comments