Recursive Launches Flint, a Laptop Coding Model That Was Mostly Built by AI

Flint is a coding model that runs on your laptop

Recursive, the AI lab led by former Salesforce chief scientist Richard Socher, released its first model this week. It is called Flint. The Flint coding model writes code and is small enough to run on a laptop. The model is free under an Apache 2.0 license, so anyone can download, use, and modify it. That alone would be notable. What makes Flint a story is Recursive's claim that the model was largely built by the company's own automated research system, the Eureka Machine, which proposed ideas, ran experiments, and iterated on results across pretraining, post-training, and inference. A small team went from scratch to a released model in under four months.

The Eureka Machine did most of the work

Recursive's pitch is recursion in the literal sense: AI that improves the process of building AI. The Eureka Machine contributed to nearly every development step, according to the company. Its contributions include a sparser model shape, low-bit training that shrunk Flint by 71% with almost no quality loss, and custom kernels for Apple Silicon and Nvidia's RTX Spark, including a "Metal megakernel" that fuses the decode forward pass.

Flint is a mixture-of-experts model with 39.8 billion total parameters and about 2.5 billion active per token. The Hugging Face model card lists a 262,144-token context configuration; Recursive says up to 512k is supported, with 256k recommended. Recursive also released "rec," a lightweight inference engine for M4 Pro and M5 Pro/Max chips. Flint also works with the usual local tooling: llama.cpp, Ollama, LM Studio, and vLLM.

Socher's summary on X was characteristically blunt: "Private, unmetered intelligence on your own machine."

How fast is it, really?

Speed is Flint's main selling point, and Recursive's numbers are specific. On a MacBook Pro with an M4 Pro chip, the company says Flint generates about 131 tokens per second, compared with 58 to 69 for Google's Gemma 4 and Nvidia's own Nemotron 3.5 Lightning on the same machine. On Nvidia RTX Spark hardware, it claims over 150 tokens per second.

Those are company figures, not independent benchmarks, so take them with the usual grain of salt. But the direction is what matters: this is a model tuned for the chips people actually own, not a datacenter card. Wired's Will Knight tested Flint on a 48GB Mac, asking it to build a to-do app and a web-research tool, and found it handled requests where smaller models errored or lagged.

Flint coding model benchmarks: the fine print

Recursive reports 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, and 54.9% on Terminal-Bench 2.1. It says Flint beat every sub-50B model on SWE-bench Verified except Alibaba's Qwen3.8 27B, which edged it on the company's six-test average.

Two caveats are worth stating plainly. First, these are vendor-reported numbers on benchmarks that vendors increasingly optimize for. Second, Recursive itself admits Flint may need more guidance than larger closed-source coding assistants. A model that runs locally is not the same as a model that replaces a frontier API.

If you follow the open-weight scene, Flint joins a busy month. Nvidia open-sourced five retrieval models built for smarter AI agents, and Inception's Mercury 3 pushed a diffusion-based architecture to over 1,000 tokens per second. The trend is unmistakable: smaller, specialized models are punching far above their weight.

Why local coding models matter

The case for Flint is not that it beats the biggest models. It is that coding assistance is becoming infrastructure, and infrastructure you run yourself has different economics: no per-token bill, no data leaving your machine, no outage on someone else's status page taking your workflow down with it.

The more interesting story is the Eureka Machine. If a small team can ship a competitive coding model in four months because an automated research system did the experimental grunt work, the moat around model development gets thinner. Every lab is trying some version of this; Recursive is just saying it out loud, with a release to point at. The company describes its mission as building "recursively self-improving superintelligence to automate knowledge discovery," and calls Flint its starting point. Starting points are easy to announce. The next releases will tell us whether the machine is as good as its marketing.