Recursive releases Flint, a free laptop coding model built largely with AI

Recursive has released Flint, a free coding model designed to run on a laptop, and says much of the model’s development was handled with help from its own AI research system. The launch matters because it points to a lower-cost path for developers who want local coding help without sending code to a cloud service.

The model is also an early public test of Recursive’s broader thesis: AI systems that help build better AI systems. For freelancers, small agencies and technical teams watching coding costs, Flint is less about replacing premium assistants immediately and more about testing what local AI can do when privacy, offline access and usage fees matter.

Flint brings coding assistance onto the laptop

According to Recursive, Flint can run on a user’s own computer rather than through company-run cloud servers. That means developers can work offline, keep code on their device and avoid cloud AI services that charge based on the amount of text processed.

Recursive says Flint is available to download for free on Hugging Face under the Apache 2.0 license, which allows free use and modification. That combination makes it especially relevant for developers who want to experiment with local coding assistants inside their own workflows, including client work where code privacy is a practical concern.

For people already using hosted coding tools such as GitHub Copilot, Cursor or Replit, Flint sits in a different part of the market. Those services are built around connected development environments and cloud-backed AI assistance, while Flint’s pitch is local execution and control. The tradeoff, by Recursive’s own framing, is that users may need to guide Flint more carefully than they would a large paid assistant.

Built with the Eureka Machine

Recursive says Flint was developed with major help from its automated research system, which it calls the Eureka Machine. The company says that system helped with nearly every stage of the model’s creation, including pretraining, post-training and inference.

The company told TechCrunch that this allowed a small team to build Flint from scratch in under four months. Recursive describes the Eureka Machine as a system that generates ideas, runs experiments and uses results to choose the next steps.

Recursive says Eureka contributed several technical pieces to Flint, including:

• a sparser model structure;
• low-bit training that reduced Flint’s size by 71% with little quality loss, according to the company;
• custom kernels designed to speed up the computations behind the model.

The company did not say how much researcher time the Eureka Machine saved. But the release fits a broader industry move toward using AI agents in research and engineering work; TechCrunch notes that OpenAI recently said its AI agents were logging 3.1 workdays for every human researcher workday by mid-August.

Performance claims focus on small, local models

Flint is a mixture-of-experts model, meaning it uses only part of its total capacity for a given task. Recursive says the model has 40 billion parameters in total, but activates 2.5 billion at a time.

That design is central to its laptop pitch. TechCrunch contrasts Flint with Meta’s Muse Spark Open, an open model for Macs released the previous week, which the report says uses 816 billion parameters and requires at least a terabyte of memory.

Recursive also published speed claims. In the company’s tests, Flint generated about 131 tokens per second on a MacBook Pro with Apple’s M4 Pro chip. TechCrunch reports that Recursive compared this with 58 to 69 tokens per second for similar models on the same machine, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning. On Nvidia’s RTX Spark, Flint exceeded 150 tokens per second, according to the company.

On SWE-Bench Verified, which measures AI performance on real-world software bug fixes, Recursive says Flint beat other models below 50 billion parameters. On the company’s own average across six tests, TechCrunch reports that only Alibaba’s Qwen3.8 27B scored higher.

Why small teams should pay attention

Recursive is careful not to position Flint as a full replacement for the largest paid coding assistants. The company says users should break bigger jobs into smaller steps and expect to provide more direction.

That limitation matters. A freelancer using AI to scaffold a small tool, inspect a script or prototype an internal app may find a local model useful if it performs reliably enough. A team expecting a coding assistant to handle complex, multi-file work with minimal supervision may still prefer larger hosted systems.

Early outside testing was encouraging, according to TechCrunch. Wired’s Will Knight reportedly used Flint on a Mac with 48GB of memory to build several small programs, including a to-do app and a web research tool, and found that Flint handled his requests successfully while smaller coding models he had tried were often slower or more error-prone than Claude or Codex.

The larger question is whether Recursive can keep improving its own research loop. The startup, founded last year and led by CEO Richard Socher, emerged from stealth in May with $650 million in funding at a $4.65 billion valuation from backers including GV, Nvidia and AMD Ventures. It also previously told TechCrunch that a $410 million compute deal with Amazon Web Services in June was focused less on headcount than on “agent count.”

For now, Flint is a practical artifact of that strategy: a free local coding model that developers can test today, and a public signal that Recursive wants its AI research assistant to help build the next generation of models faster.