Recursive Launches Flint, an AI-Built Laptop Coding Model

Recursive released its first model, Flint, on Wednesday. The startup, founded last year, says AI provided much of the assistance used to develop the free laptop coding model. Recursive says its automated research system, the Eureka Machine, assisted with almost every development stage. Recursive says that assistance enabled a small team to create Flint from scratch in under four months.

Flint provides an early public test of recursive self-improvement, in which AI creates better AI that subsequently develops still better systems. Some researchers fear that self-improving AI could escape human control. However, Recursive currently uses a more limited approach: a human-supervised AI research assistant that produced a model suitable for a laptop. Recursive aims to develop AI capable of improving itself.

Recursive emerged from stealth in May with $650 million in funding and a valuation of $4.65 billion. Recursive's backers include GV, Nvidia and AMD Ventures. CEO Richard Socher, previously Salesforce's chief scientist, leads Recursive.

Flint operates on developers' computers instead of company cloud servers, allowing offline use, code privacy and avoidance of text-processing usage fees. Developers can download Flint free from Hugging Face under the Apache 2.0 open-source license, which permits free use and modification. Flint uses a mixture-of-experts architecture that activates just 2.5 billion of the model's 40 billion parameters for a given task. That architecture makes Flint suitable for an ordinary laptop and much smaller than Meta's Muse Spark Open. Meta released Muse Spark Open last week as an open model designed for Macs. Muse Spark Open has 816 billion parameters and requires at least a terabyte of memory.

Recursive's tests showed Flint producing about 131 tokens, or text chunks, per second on a MacBook Pro equipped with Apple's M4 Pro chip. Recursive's tests put comparable models, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, at 58 to 69 tokens per second on that computer, roughly half Flint's speed. Flint exceeded 150 tokens per second on Nvidia's RTX Spark. Flint beat other models with fewer than 50 billion parameters on SWE-Bench Verified, which tests the ability to fix real-world software bugs. Recursive's average across six tests placed only Alibaba's Qwen3.8 27B ahead of Flint.

Recursive acknowledged that Flint would require more guidance than major paid coding assistants and recommended dividing large tasks into smaller steps. Early reviews were favorable. Wired's Will Knight used Flint on a Mac with 48GB of memory to create several small programs, including a to-do list application and a web research tool. Knight found that smaller coding models often produced errors or operated slowly compared with Claude or Codex. However, Knight found that Flint fulfilled every coding request.

The Eureka Machine generates ideas, conducts experiments and selects subsequent experiments using the results. Recursive's launch video says Eureka contributed across almost the entire Flint development process, including pre-training, post-training and inference. Recursive said Eureka contributed a sparser model architecture. Recursive said Eureka also contributed low-bit training that reduced Flint's size by 71% while causing almost no quality loss. Recursive said Eureka contributed custom kernels, specialized programs that accelerate underlying computations. Recursive did not disclose how much researcher time Eureka saved. OpenAI said last month that, by mid-August, its AI agents were recording 3.1 workdays for each workday completed by a human researcher.

If Recursive's approach succeeds, successive models should handle longer periods of work without human assistance. Recursive previously described the company's $410 million compute agreement with Amazon Web Services in June as "less about headcount and more about agent count." Socher told Wired that the ultimate goal is for Eureka to achieve full recursive self-improvement, with humans supervising and able to halt the process if problems arise. Recursive says successive models should improve Eureka's research capabilities and help Eureka develop the following model faster.