Recursive Releases Flint, a Free 40-Billion-Parameter Coding Model That Runs on a Laptop

Recursive released Flint on Wednesday, October 21, 2026. It is the company's first model, a free coding model that runs on a laptop, and it can be downloaded from Hugging Face under the Apache 2.0 open-source license. TechCrunch reported the launch the same day.

Flint is a mixture-of-experts (MoE) model with 40 billion total parameters, of which 2.5 billion are active at any one time. Think of it as a large team of specialists where only a few are consulted for each token, the chunk of text the model reads and writes. Sparse activation is what makes a model of this size practical on consumer hardware, because per-token compute tracks the active parameters rather than the full count. The weights still have to fit in memory. That is the constraint separating Flint from models that need a server rack.

On speed, Recursive says that in its tests Flint generated about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip. That is roughly twice the 58 to 69 tokens per second the company measured for similar models, including Gemma 4 and Nemotron 3.5 Lightning. On Nvidia's RTX Spark, Recursive reports throughput above 150 tokens per second.

On SWE-Bench Verified, a benchmark that tests whether a model can resolve real GitHub issues, Flint outperformed other models with fewer than 50 billion parameters. On Recursive's six-test average, only Alibaba's Qwen3.8 27B scored higher. The company's Introducing Flint post describes Flint as the best-performing mixture-of-experts model under 50 billion total parameters on agentic coding benchmarks, and says it rivals the strongest dense models of similar size.

Wired's Will Knight ran Flint on a 48GB Mac and used it to build several small programs. He reported that Flint handled each of his coding requests, while smaller coding models he had tried often produced errors or responded slowly.

The development process is the other element Recursive is emphasizing. The company says its automated research system, called the Eureka Machine, helped with nearly every step of Flint's development, allowing a small team to build the model from scratch in under four months. Recursive's First Steps Toward Automated AI Research post describes a system that automates the research loop for a target objective. The company's updates page had already described Flint as a local model with the fastest inference and strongest coding performance in its class.

Recursive emerged from stealth in May 2026 with $650 million in funding at a $4.65 billion valuation. Backers include GV, Nvidia, and AMD Ventures. Richard Socher, Salesforce's former chief scientist, is the CEO.

Flint arrived a week after a model at the opposite end of the local-deployment spectrum. Meta's Muse Spark Open, an open-weight model built for Macs, was released the week before Flint. It uses 816 billion parameters and needs at least a terabyte of memory.

In my view, the two releases define the range of what "local" now means for developers. A 2.5-billion-active-parameter model that runs on a laptop and an 816-billion-parameter open-weight model that needs a workstation with a terabyte of memory serve different users with different budgets. The broader context here is that both are open to download, which lets teams compare them directly on their own code rather than relying on a vendor's benchmark table.

The claim most worth watching is the one about process: an automated research loop building a competitive model in under four months. If that method can be repeated, the cost and time required to produce a usable specialized model would fall. If it cannot, Flint still stands as a model developers can test directly.

For developers, the practical gain is simple to state. A capable coding model that runs on hardware they already own lets them keep source code and prompts on their own machines, at no download cost, and test it against their own repositories before deciding whether it earns a place in their workflow.