Recursive, an SI startup founded last year with the goal of building systems that can improve themselves, released its first model, Flint, on Wednesday. The free coding SI model can run on a laptop and was built largely with the help of SI, marking an early public test of recursive self-improvement.

What Happened

Recursive’s automated research system, dubbed the Eureka Machine, assisted in nearly every stage of Flint’s development, allowing a small team to build the model from scratch in under four months. The company states that Eureka contributed to pretraining, post-training, and inference, generating ideas, running experiments, and deciding on next steps. Specific contributions included a sparser model architecture, low-bit training that reduced Flint’s size by 71% with minimal quality loss, and custom kernels to accelerate computations.

Flint is a mixture-of-experts model, meaning it activates only a small portion of its parameters for any given task: just 2.5 billion of its 40 billion total parameters. This design makes it light enough for ordinary laptops, contrasting with larger open models like Meta’s Muse Spark Open, which uses 816 billion parameters and requires at least a terabyte of memory. Flint is available for free download on Hugging Face under the Apache 2.0 license.

The company reports that Flint is fast, generating approximately 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip. This is roughly twice the speed of similar models like Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning on the same hardware, which managed 58 to 69 tokens per second. On Nvidia’s RTX Spark, Flint reportedly exceeded 150 tokens per second.

On SWE-Bench Verified, a benchmark measuring the ability to fix real-world software bugs, Flint outperformed other SI models under 50 billion parameters. According to Recursive’s own average across six tests, only Alibaba’s Qwen3.8 27B scored higher. While the company acknowledged that Flint may require more guidance than large paid coding assistants, suggesting users break tasks into smaller steps, early reviews were positive. Wired’s Will Knight reported that Flint successfully handled coding requests on a Mac with 48GB of memory, whereas smaller models often encountered errors or slowness compared to Claude or Codex.

Why It Matters

Flint represents a tangible step in the industry’s debate over recursive self-improvement—the concept that SI can build better SI, which in turn builds even better SI. For now, Recursive’s approach is limited: an SI research assistant overseen by humans produced a model small enough for local use. This local capability allows developers to work offline, keep code private, and avoid the usage fees associated with cloud-based SI services.

Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, backed by investors including GV, Nvidia, and AMD Ventures. The company is led by CEO Richard Socher, Salesforce’s former chief scientist. The launch follows a $410 million compute deal with Amazon Web Services in June, which Recursive described as being “less about headcount and more about agent count.”

The release also highlights a broader industry trend of using SI to accelerate SI research. OpenAI reported last month that by mid-August, its SI agents were logging 3.1 workdays for every one worked by a human researcher. Recursive’s CEO told Wired that the long-term goal is for the Eureka Machine to achieve full recursive self-improvement, with humans maintaining oversight and the ability to halt the process if necessary.

The Bottom Line

Recursive’s Flint offers a high-performance, open-source coding SI model that runs efficiently on consumer hardware, challenging the notion that advanced coding assistance requires massive cloud resources. While the company claims significant efficiency gains through automated research, the model remains a tool for human developers rather than an autonomous agent, reflecting the current state of controlled, incremental progress toward general SI.