Recursive Flint: Open-Weight Local Coding Model

Recursive, the start-up founded to build AI that improves itself, has released its first model, and it is small enough to run on a well-equipped laptop. Flint is a coding model with 40 billion parameters, of which only 2.5 billion are active for any given token. Its weights are on Hugging Face under the Apache 2.0 licence, so developers can download it, run it offline and adapt it, the company says in its launch post. Wired dated the launch to 21 October.

The more unusual claim is how it was made. Recursive says a “small team of people and agents built it from scratch in under four months”, with its automated research system, the Eureka Machine, helping on “nearly every part” of the work, from data and training recipes to the code that runs it.

What Recursive says Flint can do

Flint uses a mixture-of-experts design, in which only some of the model’s parts switch on for each token. That keeps it fast. According to Recursive, it leads all such models with fewer than 50 billion total parameters on three agentic coding benchmarks, SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified, and rivals dense models of 25 to 40 billion parameters “while running more than five times faster”.

On speed, the company says its own inference engine gets more than 125 tokens a second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark machine at short contexts, which it calls “up to twice as fast as comparable models”. Flint accepts up to 512,000 tokens of context, though Recursive recommends staying within 256,000. All of these figures are Recursive’s own, and no outside benchmark run has been published yet.

The company is candid about limits. Flint “may need more guidance than the larger closed-source coding models you’re used to”, it says, and works best when big jobs are broken into smaller tasks with explicit instructions.

An AI that helped tune the AI

Recursive gives a few examples of what the Eureka Machine did. It searched for an architecture, starting from a 35-billion-parameter baseline with 3 billion active, and settled on the 40/2.5 split as the best mix of speed and coding performance among the configurations it tried. After training, it went against common practice by squeezing the model’s output layer to 4 bits, then added a distillation step that Recursive says nearly halved training steps against its baseline. The company says the final compressed NVFP4 version scores within 0.7 points of the full 16-bit model. The system also wrote specialised code for Apple Silicon and RTX Spark, including a single fused “megakernel” for M4 and M5 chips.

Flint works with common local tools, including llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang. Recursive is also releasing rec, a small inference engine tuned for Flint on M4 Pro and M5 Pro and Max chips.

A test case for self-improving AI

Wired’s Will Knight, who tested Flint on a Mac with 48GB of memory, wrote that it “nailed my coding requests every time”, building a command-line to-do list, a research-scouring program and a simple website. That is one reporter’s experience, not a benchmark.

The bigger argument is about recursive self-improvement, in which AI designs its own successors. Some researchers fear it could slip out of human control, and Wired notes that several bills in Congress propose to ban it. Richard Socher, Recursive’s chief executive and cofounder, told Wired that Flint was built with a system “incapable of escaping human control”, and that any fully self-improving version would still be overseen by people who could stop it. “It’s the only way this very small team can produce a model that’s actually better than any other model in its class,” he said.

Recursive came out of stealth in May with $650m in funding, according to coverage it links on its own site. Flint is its starting point. Whether the next model arrives faster, as the company bets, is the real test of the Eureka Machine.