Recursive Flint: A 40B Coding Model You Can Run on a 48GB Mac

Recursive Flint is a new open-weight coding model that Recursive Superintelligence released on October 21, 2026. It has about 40 billion parameters but only activates 2.5 billion per token, so a 48 GB Mac can run it locally.

• Released October 21, 2026, as the company’s first model.
• Apache 2.0 license, so you can use it commercially.
• 39.8B total parameters, about 2.5B active per token.
• Text-only, with a 262K-token context window and tool calling.
• Every benchmark number so far comes from Recursive itself.

What Recursive Flint actually is

Flint is a mixture-of-experts model built for coding agents. It’s meant to explore a codebase, fix bugs, add features and write tests.

The trick is in the routing. Each layer holds 256 experts, but the router picks only 6 per token (plus one shared expert). That is why a 40B model can generate text at the cost of a much smaller one.

It handles text only, so you can’t feed it screenshots. Rivals such as Gemma 4 and Qwen3.6 accept images, as MarkTechPost points out.

What hardware you need to run it locally

Recursive ships three formats, and the memory numbers matter more than the parameter count. Here is how they compare.

For most people, the first row is the one that counts. A MacBook Pro with 48 GB can hold the 4-bit version, and Recursive reports over 125 tokens per second on an M4 Pro. That figure is self-reported and measured at short context, so expect slower speeds on long projects.

Flint also works with llama.cpp, Ollama, LM Studio, Unsloth, vLLM and SGLang. If you’d rather not download anything, there’s a free OpenRouter endpoint with tool calling.

How good is it? Read the benchmarks carefully

Recursive’s headline number is 89.7% on SWE-bench Pro v2. Here are the main scores it published.

This is where I’d slow down. Recursive measured “SWE-bench Pro v2” itself, and MarkTechPost notes it isn’t comparable to the public SWE-Bench Pro. The more familiar SWE-bench Verified score is 74.2%.

The same report says Flint trails the strongest dense models. Dense Qwen 3.8, for example, scores 73.0% on Terminal-Bench 2.1 against Flint’s 54.9%. Wait for independent tests before you trust the top-line figure.

A practical tip: start with the free OpenRouter endpoint to see whether Flint suits your style of coding. Only download the 22 GB file once you know you’ll use it, because that saves disk space and a long wait on a slow connection.

The 4-bit trick and the “Eureka Machine”

Recursive says an automated research system it calls the Eureka Machine helped design the architecture, the quantization and the kernels. That’s a bold claim, and we only have the company’s word for it.

One unusual choice is that the team quantized the output layer to 4 bits, which most teams avoid. They then used the 16-bit model as a teacher to win back accuracy. The result: the NVFP4 build lands within 0.7 points of full precision on the reported tests.

If you’ve followed other open releases, such as our look at DeepSeek DeepInfer going open source, you’ll know the pattern. The weights are open, but the full recipe often isn’t.

Should you swap your coding assistant for Flint?

Probably not yet, but it’s worth a weekend test. A free, private model that runs offline is useful for code you can’t send to a cloud service.

It also arrives while big labs hold back their largest models. Our post on Mistral Large 4 and its missing open weights shows how rare a fully downloadable release can be.

Try it in a side project first. Compare it with your current tool on three real tasks, such as a bug fix, a small feature and a test file.

Frequently Asked Questions

Is Recursive Flint free to use?

Yes. The weights use the Apache 2.0 license, which allows commercial use. A free hosted endpoint is also available on OpenRouter.

Can I run Flint on a laptop?

Yes, if it has enough memory. The 4-bit Apple Silicon build is about 22.6 GB, and Recursive recommends 48 GB of memory.

Is Flint better than Qwen or Gemma?

Not clearly. It scores well on Recursive’s own tests, but the same report says it trails the strongest dense models like Qwen 3.8 on Terminal-Bench 2.1.

Does Flint understand images?

No. It’s text-only, so it can’t read screenshots or diagrams.

Our take

Flint is the most interesting local coding model to land this month, mainly because 48 GB of memory is within reach of a serious developer’s laptop. I’d treat the 89.7% score as marketing until outside testers confirm it. Download the 4-bit build, run it on your own code, and let your real projects decide.