Recursive launches Flint, a coding model for laptops

By VK

Published 22 Oct 2026, 10:39 UTC

Reporting from TechCrunch, WIRED

Richard Socher's Recursive says its automated research system built most of Flint, a free 40B open-weight coding model that runs locally.

What it means for founders

A local coding model with no token bill. Teams that cannot send code to a cloud API, or that watch inference spend, have a free option worth trialling on a well-specced Mac. Pull it through Ollama, keep tasks narrow, and compare it with Claude Code before moving real work over.
You can build on it. Apache 2.0 permits free use and modification, so a fine-tuned Flint could ship inside a product with no per-call fees. Recursive's site lists a Model Terms page, so check whether it adds anything to the license first.
Small models may get cheaper to make. If AI-driven research let a handful of people build a competitive model in under four months, and others copy the method, specialized open models could arrive faster. Recursive says more local models will come from the same process. Watch the gap before its next release: Recursive's stated bet is that each stronger model improves the Eureka Machine's research, so a clearly shorter cycle would be one signal the loop works.

The story

Recursive, a startup run by former Salesforce chief scientist Richard Socher that is trying to build self-improving AI, released its first model on Wednesday, October 21. Flint is a laptop-sized coding model that, the company says, was largely designed and tuned by its automated research system, the Eureka Machine. The weights are downloadable from Hugging Face under an Apache 2.0 license.

What Recursive shipped

Flint uses a mixture-of-experts design: 40 billion parameters overall, with 2.5 billion switched on per token. It works with Ollama, llama.cpp, vLLM and other runtimes, and Recursive released an inference engine, rec, tuned for Apple's M4 Pro, M5 Pro and M5 Max chips.

Using its own inference engine at short contexts, Recursive clocked Flint above 125 tokens a second on Apple's M4 Pro MacBook Pro, and above 150 on Nvidia's 64GB RTX Spark. Recursive calls that up to twice the pace of comparable models, and claims Flint leads mixture-of-experts models under 50 billion parameters on SWE-Bench Verified, SWE-Bench Pro v2 and Terminal-Bench 2.1. TechCrunch notes that Alibaba's Qwen3.8 27B still beat it on the company's six-test average. Recursive warns that Flint may want more steering than the big closed-source assistants, and its advice is to feed it small, clearly specified subtasks.

Will Knight of WIRED ran it on a Mac carrying 48GB of memory. Flint got every coding request right, he wrote, including a terminal to-do tracker and a web research tool, where models its size had tended to be buggy or slow for him.

How much of Flint did AI build

Socher told WIRED that "nearly every aspect of the entire model pipeline was actually done and improved upon by AI." Recursive says the Eureka Machine proposed and tested ideas across data, training recipes, architecture search, 4-bit quantization and hardware-specific kernels, letting a few researchers create Flint from nothing in less than four months. People still supervised the work, and no figure for the hours it saved has been given.

Investors such as GV, Nvidia and AMD Ventures put $650 million into Recursive when it came out of stealth in May, valuing it at $4.65 billion. Recursive's goal is full recursive self-improvement that humans can halt, a practice several congressional bills would outlaw, according to WIRED.

What we don't know yet

How the work on Flint split between the Eureka Machine and its human researchers. Recursive has not published a breakdown.
Whether the speed and benchmark figures survive independent testing.
How Recursive will make money from a free model.