Flint: a free coding model you can run on your own laptop. What is claimed and what is not proven
Recursive released Flint on 21 October: a free, Apache 2.0 mixture-of-experts coding model built to run locally. The speed and benchmark numbers come from the company itself, so here is what is reported, what was tested by others, and what to check before relying on it.
On 21 October 2026 the AI startup Recursive released Flint, a free coding model designed to run on your own computer. It is on Hugging Face under the Apache 2.0 licence, which allows commercial use. For developers who want offline use or do not want to send code to a cloud service, that makes it worth a look.
Almost everything below comes from one TechCrunch report, and nearly all performance numbers come from Recursive itself. We have not run the model, and we separate claims from checks.
What is reported
- Maker: Recursive, a startup founded in 2025 and led by CEO Richard Socher. TechCrunch says it came out of stealth in May 2026 with $650 million at a $4.65 billion valuation.
- Design: a mixture-of-experts model with 40 billion parameters in total, of which about 2.5 billion are active for each task. That is why it can run faster than its size suggests.
- Licence and access: free, downloadable from Hugging Face, Apache 2.0.
- Hardware: built to run locally, offline, with no cloud fees. TechCrunch gives no minimum memory requirement. It mentions only a Mac with 48 GB of memory used in a Wired review.
Claims from the company (not independently verified)
| Claim | Source |
|---|---|
| About 131 tokens per second on a MacBook Pro with an M4 Pro chip | Recursive's tests |
| Over 150 tokens per second on an Nvidia RTX Spark | Recursive's tests |
| Beats other models under 50B parameters on SWE-Bench Verified | Recursive |
| On Recursive's own six-test average, only Alibaba's Qwen3.8 27B scored higher | Recursive |
| Built in under four months by a small team, with an automated research system ("Eureka Machine") helping with almost every step | Recursive / its CEO |
| Low-bit training shrank the model by 71% with "almost no loss in quality" | CEO's posts |
TechCrunch does not give the exact scores or say whether anyone outside Recursive ran these tests. The only outside view mentioned is one Wired reviewer, who said the model handled his requests well. That is a single anecdote.
What Recursive admits
Flint needs more guidance than large paid coding assistants. Recursive suggests breaking work into smaller steps. Expect to prompt it more carefully than a top cloud model.
Should you try it?
It is a reasonable thing to test, with these cautions:
- Check memory first. The model card on Hugging Face should list the file sizes. A 40B model, even compressed, needs far more than a typical 8 to 16 GB laptop, and the company gives no minimum.
- Test it on your own code, not on benchmark scores. Try a bug fix, a refactor and a unit test in a repository you know.
- Compare speed on your machine. The 131 tokens per second figure is from one chip and one setup.
- Keep sensitive code local only if you confirm the tool you use to run it does not send data elsewhere.
- Wait for independent benchmarks before choosing it for a team.
The aim of full recursive self-improvement, with human oversight, is described as a goal, not something shown to work.
Update log
- 25 October 2026: first published from TechCrunch's 21 October report. One figure in that report differs from its linked headline (Recursive's AWS compute deal is given as $410 million in one place and $400 million in a headline), so we leave it out. We will update when independent tests appear.
Cover photo: KK IN HK, CC BY-SA 4.0, via Wikimedia Commons.
