Recursive Superintelligence released Flint on October 21, a coding model with 40 billion total parameters that runs directly on high-performance consumer laptops. What makes this story different from hundreds of similar releases is that Recursive claims nearly every step of Flint's development was carried out by the company's own automated research system — the Eureka Machine — rather than by human engineers.
To run on the user's machine, Flint activates only 2.5 billion of its 40 billion parameters per token, using a mixture-of-experts (MoE) architecture. In practice, this means the model uses a small slice of itself for each task. The result is a model light enough for a consumer laptop but powerful enough to compete on agentic coding benchmarks against much larger closed models.
In the company's tests on SWE-Bench Verified, Flint outperformed all other MoE models with fewer than 50 billion parameters on the market. In Recursive's own benchmark suite — a set of six tests including Terminal-Bench 2.1 and SWE-Bench Pro v2 — only one model, Alibaba's Qwen3.8 27B, scored higher. Speed is also a key differentiator: on a MacBook Pro with an M4 Pro chip, the model generated approximately 131 tokens per second, roughly double the rate of similar models like Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning.
The speed is not an accidental byproduct. Recursive optimized the entire inference stack. The Eureka Machine quantized the output layer (lm_head) to four bits — against usual practice, which keeps output layers at higher precision — and recovered accuracy with offline distillation, training the quantized model on pre-sampled responses from the original 16-bit model. The published losses were minimal: from 74.2 to 74.1 on SWE-Bench Verified, and from 54.9 to 53.7 on Terminal-Bench 2.1. That is the difference between a 40-billion-parameter model that needs a datacenter GPU and one that fits in a high-end laptop's memory.
But Recursive's boldest claim is not about benchmarks or latency. It is about process. According to Richard Socher, CEO of the company and former chief scientist at Salesforce, "nearly every aspect of the entire model pipeline was actually done and improved upon by AI." The Eureka Machine generates ideas, runs experiments, reads results, and decides which attempt to try next — and according to the company, it contributed to practically every stage of building Flint, from architecture and training recipes to infrastructure and inference optimization. A small team built the model from scratch in under four months.
The concept of AI building better AI — the so-called recursive self-improvement — is already one of the most debated topics in the industry. Some researchers fear an auto-improving system could slip from human control. For now, Recursive's version is limited: it is an AI research assistant supervised by people that produced a model small enough to run on laptops. Socher told Wired that the long-term goal is for the Eureka Machine to achieve full recursive self-improvement, "with humans overseeing the process and able to stop it if something goes wrong."
Flint is free to download on Hugging Face under the Apache 2.0 license, which permits unrestricted use, modification, and distribution. Recursive also released a lightweight inference engine optimized for M4 Pro and M5 chips, compatible with llama.cpp, Ollama, LM Studio, vLLM, and SGLang. The model supports context windows of up to 512,000 tokens, though the company recommends staying below 256,000 for best results.
Recursive emerged from stealth in May with US$650 million in funding from investors including GV, Nvidia, and AMD Ventures, at a US$4.65 billion valuation. It also secured a US$410 million compute deal with Amazon Web Services.
The company openly acknowledges that Flint needs more guidance than large closed coding assistants and recommends breaking complex tasks into smaller subtasks. This is an important distinction: Flint does not replace Claude or Codex for heavy refactoring. For interactive work — code completions, test generation, small fixes — latency is the deciding factor, and that is exactly the niche Flint is positioned for.
What Recursive achieved with Flint is not just a competitive model. It is a prototype of how AI research might be done in the future: fewer meetings and more closed loops of hypothesis-experiment-result. What remains to be seen is whether the Eureka Machine can, in its next cycles, create a model so capable that it no longer needs human supervisors — or whether the real test of the model is that it becomes so useful that developers choose not to stop it.
Sources: TechCrunch, RITS Shanghai NYU, Dealroom
✓ Independent sources cross-checked and verified before publishing
