Recursive Flint: a 40B coding model that runs at 131 tokens a second on a MacBook, and its makers say AI built most of it
Recursive, the lab Richard Socher runs, released Flint on Hugging Face on October 21 under Apache 2.0. It is a mixture-of-experts coding model with 39.8 billion total parameters and about 2.5 billion active per token, 256 routed experts with 6 chosen at a time, and a 262,144-token context. TechCrunch measured the pitch: about 131 tokens per second on an M4 Pro MacBook Pro, roughly twice what Gemma 4 and Nemotron 3.5 Lightning manage on the same laptop. Link: https://huggingface.co/RecursiveSI/flint
Self-reported scores: 54.9% on Terminal-Bench 2.1, 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, averaged over at least five attempts with off-the-shelf harnesses (Terminus 2 and mini-swe-agent). Recursive's own table is honest that the dense Qwen3.8 27B still beats it on the six-test average. Flint's claim is narrower: best among MoE models under 50 billion total parameters, and fast enough that a local coding agent stops feeling like a punishment.
The more interesting claim is how it was made. Recursive says a small team of people and agents built it from scratch in under four months, and that its automated research system, the Eureka Machine, helped develop nearly every part of the stack. TechCrunch lists the specific contributions: a sparser model shape, low-bit training that shrank Flint by 71% with almost no quality loss, and custom kernels. The NVFP4 version lands within 0.7 points of the 16-bit one. Community GGUF and MLX quants appeared within hours.
Treat the "AI built it" line as marketing until someone outside Recursive can audit the loop. But the parts credited to the machine are exactly the parts an automated search is good at: shapes, quantization recipes, kernels, things with a number to optimize. This is the first shipped model sold as the output of a research loop rather than a research team. The model is useful today. The real product being advertised is the loop.
← Back to all articles
Self-reported scores: 54.9% on Terminal-Bench 2.1, 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, averaged over at least five attempts with off-the-shelf harnesses (Terminus 2 and mini-swe-agent). Recursive's own table is honest that the dense Qwen3.8 27B still beats it on the six-test average. Flint's claim is narrower: best among MoE models under 50 billion total parameters, and fast enough that a local coding agent stops feeling like a punishment.
The more interesting claim is how it was made. Recursive says a small team of people and agents built it from scratch in under four months, and that its automated research system, the Eureka Machine, helped develop nearly every part of the stack. TechCrunch lists the specific contributions: a sparser model shape, low-bit training that shrank Flint by 71% with almost no quality loss, and custom kernels. The NVFP4 version lands within 0.7 points of the 16-bit one. Community GGUF and MLX quants appeared within hours.
Treat the "AI built it" line as marketing until someone outside Recursive can audit the loop. But the parts credited to the machine are exactly the parts an automated search is good at: shapes, quantization recipes, kernels, things with a number to optimize. This is the first shipped model sold as the output of a research loop rather than a research team. The model is useful today. The real product being advertised is the loop.
Comments