Recursive launches Flint, a laptop coding model largely built by AI — the first public test of recursive self-improvement

• Who / what: Recursive released Flint, its first model — an open-weight coding model built to run locally on a laptop.
• When / how announced: October 21, 2026, announced by CEO Richard Socher in a post on X.
• Specs: mixture-of-experts, about 40 billion parameters total with roughly 2.5 billion active per token; documented context window of 262,144 tokens; free on Hugging Face under the Apache 2.0 license.
• Speed (company-tested): about 131 tokens per second on an M4 Pro MacBook Pro — roughly double Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning on the same machine; over 150 tokens per second on Nvidia's RTX Spark.
• Benchmarks (company-reported, not independently verified): 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, 54.9% on Terminal-Bench 2.1; beats other models under 50 billion parameters on SWE-Bench Verified, per Recursive.
• The build: Recursive's Eureka Machine automated research system contributed a sparser model shape, low-bit training that cut the model size by 71% with almost no quality loss, and custom kernels; a small team went from scratch to finished model in under four months.
• Money: Recursive emerged from stealth in May with $650 million at a $4.65 billion valuation (backers include GV, Nvidia and AMD Ventures) and holds a $410 million AWS compute deal.

Recursive, the AI startup led by former Salesforce chief scientist Richard Socher, has released its first model — and the model itself is the pitch. Flint, a free, open-weight coding model that runs on a laptop, was built in under four months with Recursive's automated research system, the Eureka Machine, doing most of the work, the company said on Wednesday, October 21.

What Flint is

Flint is a mixture-of-experts model for agentic coding: investigating codebases, fixing bugs, adding features and writing tests. Of its roughly 40 billion parameters, only about 2.5 billion are active for any given token, which keeps it light enough for an ordinary laptop. Its model card documents a 262,144-token context window, with reasoning and tool calls supported.

The weights are free to download on Hugging Face under the Apache 2.0 license, which allows commercial use and modification. Because it runs on the user's own machine rather than a company's cloud servers, developers can use it offline, keep their code private, and avoid the per-token usage fees that cloud AI services charge.

The numbers

Recursive says Flint is fast. In the company's tests it generated about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip — roughly twice the 58 to 69 tokens per second that similar models managed on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark it topped 150 tokens per second.

On benchmarks, Recursive reports 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, and says Flint outperformed every other model under 50 billion parameters on SWE-Bench Verified. Across the company's own average of six tests, only one model — Alibaba's Qwen3.8 27B — scored higher than Flint. These are company-run figures and have not been independently verified.

The Eureka Machine

The Eureka Machine is Recursive's automated research system: it comes up with ideas, runs experiments and uses the results to decide what to try next. According to the company's launch video, it contributed to nearly every part of building Flint, from pre-training to post-training and inference.

Recursive credits it with a sparser model shape, low-bit training that shrank Flint by 71% with almost no loss in quality, and custom kernels — specialized programs that speed up the underlying computations. That let what the company describes as a small team build the model from scratch in under four months, after Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation. The company also has a $410 million compute deal with Amazon Web Services, which it previously described as “less about headcount and more about agent count.”

The honest caveats

Recursive acknowledged that Flint would need more guidance than the big paid coding assistants, and suggested breaking large tasks into smaller steps. And there is a structural caveat worth keeping in mind: Recursive is both the vendor and the judge here — the company that built the Eureka Machine is also the one telling us how much credit it deserves. The real test is what Flint does once developers run it on their own laptops and compare it against models built the old-fashioned way.

Early hands-on impressions are positive: Wired's Will Knight used Flint on a Mac with 48GB of memory to build several small programs, including a to-do list app and a tool that scours the web for research, and found it nailed his coding requests where smaller coding models often ran into errors or were slow compared with Claude or Codex.

Why it matters

Flint is an early public test of recursive self-improvement — the idea that AI can build better AI, which in turn builds even better AI. It is one of the most debated ideas in the industry; some researchers fear self-improving AI could slip out of human control. Recursive's version is deliberately modest: an AI research assistant, overseen by people, that produced a model small enough for a laptop.

Socher's stated goal is for the Eureka Machine to reach full recursive self-improvement, still overseen by people who could stop it. If the bet pays off, each new model should make the Eureka Machine a better researcher, helping it build the next one faster — a loop Recursive is now running in public. (Primary source: Socher's announcement thread on X.)