Recursive released its first model on 21 October 2026: Flint, a free coding model that runs on a laptop. Nearly every stage of its development was assisted by the company's own research system, the Eureka Machine, letting a handful of people build it from scratch in under four months.
Key takeaways
- Mixture-of-experts: 2.5B active parameters out of 40B total
- About 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip
- Downloadable from Hugging Face under Apache 2.0, no usage fees
- Low-bit training shrank the model by 71 percent with almost no quality loss
A small model with real speed
Flint activates only a slice of its network per task and fits in an ordinary laptop's memory. Meta's Muse Spark Open, by comparison, uses 816B parameters and needs at least a terabyte of memory.
On the same MacBook Pro, competing models — Gemma 4 and Nemotron 3.5 Lightning among them — managed 58–69 tokens per second, Flint more than twice that. On Nvidia's RTX Spark it passed 150.
| Model and hardware | Tokens per second |
|---|---|
| Flint, MacBook Pro with Apple M4 Pro | about 131 |
| Gemma 4 and Nemotron 3.5 Lightning, same MacBook | 58–69 |
| Flint, Nvidia RTX Spark | over 150 |
On SWE-Bench Verified, which measures fixing real bugs, Flint beat models under 50B parameters. In the company's own average across six tests, only Alibaba's Qwen3.8 27B placed higher.
What the Eureka Machine actually did
The system forms hypotheses, runs experiments and decides what to try next. Its contributions include a sparser network shape, low-bit training and custom kernels?kernels: low-level compute routines written for a specific chip to squeeze more performance out of it.
Recursive does not say how much time this saved. OpenAI said that by mid-August its agents were logging 3.1 workdays for every single human day.
The fight over recursive self-improvement
Flint is a public test of the idea that AI builds better AI. For now it is a research assistant overseen by humans, not a loop without brakes. The company admits Flint needs more guidance than large paid coding assistants.
The topic is touchy: in September a researcher leaving Anthropic warned that AI self-improvement could end humanity within a few years. Recursive treats those fears as closer to science fiction than reality.
Why it matters
The interesting part is not the benchmark score but the ratio: a handful of researchers with a model competitive in its class. If that holds, the advantage of large labs stops being a matter of headcount. Flint does not prove full recursion, but it shows research automation with measurable effects in a product, not a slide deck.
What's next
- Recursive's goal is full self-improvement for the Eureka Machine, under human oversight
- Each new model should make the Eureka Machine a better researcher and shorten the next build
- Several bills banning recursive self-improvement are moving through Congress
