40 billion parameters, 2.5 billion active per token, and more than 125 tokens a second on an M4 Pro MacBook Pro. Those are Recursive’s numbers for Flint, the open-weight coding model it released on October 21. The company says a small team of people and AI agents built it from scratch in under four months.
I’m interested in the downloadable model. I’m more cautious about the story attached to it: AI helping build the next AI. Both deserve attention, but they need different standards of proof.
What you can download today
Recursive’s launch announcement describes Flint as a mixture-of-experts model for local coding agents. It selects a fraction of its parameters for each token instead of using the whole model every time. That helps explain the speed claim. It does not turn the download into a 2.5B model.
The official Hugging Face repository contains the unquantized BF16 checkpoint, with a download of about 80 GB. Quantized versions are separate. If you’re deciding whether Flint fits your machine, check the actual checkpoint and runtime requirements rather than reading “2.5B active” as a memory specification.
The weights are available under Apache 2.0. Recursive lists local options including Ollama, LM Studio, and llama.cpp, alongside its own optimized engine, rec. The setup documentation also covers connecting the model to coding harnesses such as Pi and OpenCode.
That is useful to me because I want model choice inside a workflow I understand. Downloadable weights give me a way to evaluate a particular version and keep using it locally, subject to the license, instead of depending entirely on a hosted endpoint.
The speed number has conditions
Recursive reports more than 125 tokens per second on the M4 Pro MacBook Pro using its optimized inference engine at short contexts. That is a specific hardware-and-software result. I would not assume the same speed on an older Mac, a different runtime, or a long coding session.
I haven’t benchmarked Flint myself. The interesting comparison for my agency work would be a small WordPress change with a known expected result: a PHP fix, a focused test, or a CSS cleanup. I’d want to see the diff and run the relevant checks before deciding whether the speed saved me any time.
Recursive is reasonably direct about the limitations. Its announcement says Flint may need more guidance than larger closed coding models and recommends breaking complex work into smaller tasks with explicit instructions. That sounds like a sensible starting point for evaluation, rather than permission to hand it an entire client build.
AI helped with the engineering
The development story is more specific than “an AI made an AI.” Recursive says its Eureka Machine contributed across the model-building process, including architecture choices, quantization recipes, and inference kernels. The company describes a system that proposes experiments, implements them, tests them, and uses the results to choose what to try next.
TechCrunch’s launch coverage makes the human oversight explicit. Flint is an early test of that research process. Its release does not establish that Recursive has achieved fully autonomous recursive self-improvement.
I find the kernel and quantization work more convincing than the grand prediction. Those are concrete engineering contributions that developers can examine and measure. Whether each generation reliably helps produce a better next generation will take more releases and evidence.
Local still needs boundaries
For client code, local inference is appealing. I can choose a setup that keeps model processing on my own hardware. But a local model connected to a shell can still make a bad change or invoke the wrong tool. Flint’s model card explicitly warns that applications need their own permissions and execution boundaries.
I’d start it in a disposable copy of a familiar project, without production credentials, and ask for one change I can verify. If it handles that consistently, I’d widen the assignment. A downloadable coding model is worth trying; access to a client’s live site is something it would still have to earn.