The key facts
- Flint is a free MoE model (40 billion parameters, 2.5 billion active) under Apache 2.0 on Hugging Face, runs offline on a laptop
- Recursive says the Eureka Machine system helped at nearly every stage of development; the model was built in under 4 months
- ~131 tokens/s on a MacBook Pro with M4 Pro; outperformed models up to 50 billion parameters on SWE-Bench Verified
The Flint launch and the idea of recursive self-improvement
AI startup Recursive, founded last year by former Salesforce chief scientist Richard Socher, has released its first model — Flint, a free coding model that runs on a laptop. According to the company, the model was created largely with the help of AI itself: the automated research system Eureka Machine was involved in nearly every stage of development, allowing a small team to build the model from scratch in under four months. Flint is an early public test of the concept of recursive self-improvement — the idea that AI can create better AI, which remains one of the most debated in the industry: some researchers fear that self-improving AI could get out of control. In its current form, however, it is a research assistant under human control that created a model compact enough for a laptop.
Funding and positioning
Recursive came out of stealth in May, raising $650 million at a $4.65 billion valuation; investors include GV, Nvidia and AMD Ventures. In June, the company signed a $410 million deal with Amazon Web Services for compute, and previously said it is 'less about headcount and more about the number of agents.' The long-term goal, Socher told Wired, is to achieve full recursive self-improvement under human oversight with the ability to stop the process in case of a failure.
Technical specifications and performance
Flint is a mixture-of-experts model: of 40 billion parameters, only 2.5 billion are active at any time, making it suitable for an ordinary laptop. For comparison, Meta's open Muse Spark Open model for Mac, released a week earlier, uses 816 billion parameters and requires at least a terabyte of memory. In Recursive's tests, Flint generated about 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip — roughly twice as fast as similar models, including Google Gemma 4 and Nvidia Nemotron 3.5 Lightning (58–69 tokens/s). On an Nvidia RTX Spark, the model exceeded 150 tokens/s. On the SWE-Bench Verified benchmark, Flint outperformed other models up to 50 billion parameters; by Recursive's averaged result across six tests, only Alibaba Qwen3.8 27B scored higher.
The Eureka Machine's contribution and the open license
According to the company, the Eureka Machine, which generates ideas, runs experiments and uses the results to choose next steps, contributed to pre-training, post-training and inference. Among the claimed results are a sparser architecture, low-bit training that reduced Flint's size by 71% with almost no loss in quality, and custom kernels to speed up computations. How much time the system saved researchers, Recursive did not say. For comparison: OpenAI said last month that by mid-August its AI agents were performing work equivalent to 3.1 working days for every day of a human researcher's work.
Limitations and early feedback
The company acknowledged that Flint will require more prompting than large paid coding assistants, and recommended breaking large tasks into smaller steps. The model is free to download on Hugging Face under the Apache 2.0 license, which allows free use and modification. Since Flint runs on the user's computer rather than in the cloud, developers can use it offline, keep their code private and avoid paying cloud AI services by the hour. Early feedback was positive: Wired journalist Will Knight used Flint on a Mac with 48 GB of memory to create several small programs, including a to-do list app and a tool for finding research on the internet. He noted that small coding models usually make mistakes or run slowly compared with Claude or Codex, but Flint, he said, handled all of his requests.
