News update: this post was researched and drafted automatically from the sources listed at the end, and published without a human edit. If you spot a mistake, write to me and I will fix it.
Recursive released Flint on Wednesday. It is an open-weight coding model that runs on a laptop and is aimed at agents. The company says the Eureka Machine, its automated research system, helped with nearly every step of building the model.
In one line: Recursive launched Flint, a free 40 billion parameter mixture-of-experts coding model with 2.5 billion active parameters that scores 74.2% on SWE-Bench Verified and runs at over 125 tokens per second on an M4 Pro MacBook Pro, built largely by AI.
What Flint is
Flint is a mixture-of-experts model. It uses only 2.5 billion of its 40 billion parameters for any given task. That makes it light enough for an ordinary laptop. Recursive has released it free on Hugging Face under the Apache 2.0 license, which allows free use and modification. Because it runs on a user's own machine, developers can work offline, keep code private, and avoid usage fees charged by cloud AI services. For comparison, Meta's Muse Spark Open, an open model built for Macs released last week, uses 816 billion parameters and needs at least a terabyte of memory. Flint is far smaller.
Benchmark results and speed
On SWE-Bench Verified, a benchmark that measures how well AI fixes real-world software bugs, Flint scored 74.2 percent, according to MarkTechPost. TechCrunch reports that Flint outperformed other models under 50 billion parameters on that benchmark. On Recursive's own average across six tests, only Alibaba's Qwen3.8 27B scored higher. Flint is also fast. In Recursive's tests, it generated about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip. Similar models managed 58 to 69 tokens per second on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark, Flint topped 150 tokens per second. TechCrunch reported that Wired's Will Knight used Flint on a Mac with 48GB of memory to build small programs, including a to-do list app and a research tool. He found that Flint nailed his coding requests every time, while smaller coding models often ran into errors or were slow compared to Claude or Codex.
The Eureka Machine behind it
Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, from GV, Nvidia, and AMD Ventures. Its CEO is Richard Socher, Salesforce's former chief scientist. The company's automated research system, the Eureka Machine, comes up with ideas, runs experiments, and uses the results to decide what to try next. Recursive says it contributed to nearly every part of building Flint, from pre-training to post-training and inference. Specific contributions include a sparser model shape, low-bit training that shrank Flint by 71 percent with almost no loss in quality, and custom kernels that speed up computations. A small team built the model from scratch in under four months. Recursive did not say how much time the Eureka Machine saved researchers. OpenAI said last month that by mid-August, its AI agents were logging 3.1 workdays for every one worked by a human researcher. Recursive's previous $410 million compute deal with Amazon Web Services in June was, in the company's words, less about headcount and more about agent count.
Where it still falls short
Recursive acknowledged that Flint would need more guidance than the big paid coding assistants. The company suggested breaking large tasks into smaller steps. That is a real limitation for anyone hoping to let an agent run a full codebase on its own. Flint is also an early public test of recursive self-improvement. For now the company's version is more limited: an AI research assistant overseen by people that produced a model small enough for a laptop. The company's longer-term goal, Socher told Wired, is for the Eureka Machine to achieve full recursive self-improvement, with humans overseeing the process and able to stop it if something goes wrong. That is not what Flint itself does today.
What I would try
Flint matters to my kind of work because it runs locally. For regulated clients who do not want code or documents leaving their premises, a local coding model with an Apache 2.0 license is worth testing. I would start small. I would use Flint for isolated tasks: writing a function, summarising a diff, or generating a test case. I would not hand it a whole client delivery yet. This touches the same problem I wrote about in A brief goes in, software comes out: shipping client products without a dev team assigned. For now, I would watch whether Flint can handle small steps in a RAG chatbot build, especially where cloud APIs are blocked. Nothing to do at scale yet, just test.