Recursive unveils Flint: a laptop coding model built largely by AI

Recursive Superintelligence, which raised $650 million with a $4.65 billion valuation in May, published its first model on Wednesday: Flint, an open-source system designed to run locally on laptops and built largely with the assistance of its own automated research engine, the Eureka Machine. The proposition is ambitious and the message is clear: Recursive believes the best way to advance artificial intelligence is with AI itself building the next version.

What Flint is

Flint is a 40-billion-parameter language model with a mixture-of-experts (MoE) architecture. For each generated token, only 2.5 billion parameters are activated, allowing the model to run on conventional hardware without relying on remote servers. In practice, this means a developer can use Flint offline, keep their code private, and avoid the usage fees that cloud AI services charge per token processed.

On Apple's MacBook Pro with the M4 Pro chip, Flint generates approximately 131 tokens per second — roughly twice as fast as comparable models such as Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, which produced between 58 and 69 tokens on the same hardware. With the Nvidia RTX Spark GPU, the model tops 150 tokens per second.

In agentic coding benchmarks, Flint positions itself as the best MoE model with fewer than 50 billion total parameters in tests including SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified. In one standard test, only the Alibaba Qwen3.8 27B model surpassed Flint.

The Eureka Machine

What makes Flint particularly noteworthy is not just what it does, but how it was created. Recursive did not use its own model to write another AI's code — instead, it used an automated research system, the Eureka Machine, to drive virtually every step of the model's development. From defining the architecture to optimizing inference kernels for Apple Silicon and Nvidia, the Eureka Machine proposed ideas, ran experiments, and used the results to decide which direction to test next.

According to the company, the Eureka Machine contributed to the model's sparse shape (an even more efficient architecture), low-bit quantization techniques that reduced the model's size by 71% with minimal quality loss, and custom kernels that accelerate projection, MoE, attention, and prompt processing operations.

Recursive's human team is described as "small" and says it built the model from scratch in under four months, with the Eureka Machine assisting at virtually every stage. The company already signed a $410 million compute deal with Amazon Web Services — and according to CEO Richard Socher, that deal was "more about agent count than headcount."

The context: recursive self-improvement

Recursive is one of several companies entering the recursive self-improvement (RSI) field, the idea that an AI can be trained to build a better AI, which in turn builds yet another. The proposition is controversial: some researchers warn that if AI begins improving itself without direct human oversight, the system could go off the rails. Recursive argues the exact opposite: human oversight is maintained at every step, and humans can stop the process at any time.

OpenAI is also using AI agents to accelerate its own research — Socher himself cited that in August, OpenAI's AI agents were accumulating 3.1 days of work for each human day. But Recursive goes further: the goal is not just to accelerate the human process, but to gradually replace the need for it.

What this means for the future

The launch of Flint represents a public test of the RSI concept. If Recursive can demonstrate that each new version of the model makes the Eureka Machine a better researcher — and that the next version is built significantly faster and more efficiently than the previous — the startup will have validated a strategy that many in the industry consider the most promising (and most dangerous) path to superintelligence.

Meanwhile, for developers who want a local code assistant without depending on expensive APIs, Flint is a free, open-source option under the Apache 2.0 license. It's available on Hugging Face and works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang.

The question that remains: if Recursive is right, and the Eureka Machine really does start building better models on its own, how many years of traditional human research will it manage to compress?

Sources: TechCrunch, Wired, Recursive

✓ Independent sources cross-checked and verified before publishing