1. Home
2. News
3. Open weights

Recursive says its automated research system helped build nearly every part of its Flint model

The 40-billion-parameter coding model runs on a laptop, and a small team built it from scratch in under four months.

By Think Facility · October 21, 2026 · Recursive Open weights Save

On October 21, 2026, Recursive released Flint, an open-weight coding model small enough to run on a laptop. It's the first model from Recursive, the startup Richard Socher runs as CEO, and the weights went up on Hugging Face under the Apache 2.0 license.

The coding scores aren't the main claim. Recursive says Flint was a practical test of its Eureka Machine, a system for automated AI research, and that the system helped develop nearly every part of how the model was built. A small team of people and agents built it from scratch in under four months.

Size

40 billion total parameters, 2.5 billion active per token

License

Apache 2.0

Speed

more than 125 tokens per second on an M4 Pro MacBook Pro

Built

from scratch in under four months

Documented context

262,144 tokens

What the Eureka Machine did

Recursive describes the system as running the whole research loop on its own: it proposes ideas, implements them, tests them, and uses the results to pick what to try next. On Flint it worked on data, optimizers and training recipes, and then on infrastructure and inference speed.

The post walks through three examples. For the architecture, the system started from a 35-billion-parameter design with 3 billion active and searched expert layouts until it landed on 40 billion with 2.5 billion active, the setup it projected would give the best mix of speed and coding skill. Then it worked on shrinking the model.

That choice makes sense for a model this sparse, since the output layer is a big share of what has to be read for every token. To win back the lost accuracy, the system added a distillation step where the small model trains on its own answers with the full-precision model as teacher. Recursive says the final 4-bit model scores within 0.7 points of the 16-bit one. Last, it wrote custom kernels for Apple Silicon and NVIDIA's RTX Spark, including one on M4 and M5 chips that fuses the whole decode step.

What the post doesn't give is a split of who did what. It says the system contributed throughout and helped the team, but there's no count of how many of Flint's design choices came from the machine rather than from people. I'd want that number before calling this automated AI research. (Recursive did publish the system's early results on test problems in June, before Flint.)

How good is it?

On its Hugging Face card, Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. It says Flint leads mixture-of-experts models under 50 billion parameters on those tests. These are its own runs, and the card warns that setups vary from lab to lab.

Josh Tobin, who writes as one of the Recursive team, framed the release around where coding happens.

LM Studio had Flint working on launch day and put a memory number on it.

One detail is worth checking yourself. The blog post says Flint supports up to 512k tokens of context and recommends staying within 256k. The model card documents 262,144 tokens. Plan around the smaller one.

What it means for people who write code

A local model changes who sees your code. Recursive's pitch is that you can work offline, keep your data on your own computer and skip API fees. Nothing gets sent to a remote server. It also says Flint needs more guidance than the big closed models and does better when work is broken into smaller pieces.

The people that pitch reaches are a large workforce. The Bureau of Labor Statistics counts about 1.7 million software developer jobs in 2025, with a median wage of $135,980 in May 2025, and projects employment of developers, testers and QA analysts to grow 10 percent from 2025 to 2035. The BLS page credits part of that demand to software being built for AI.

What a tool like the Eureka Machine does to that projection is the bigger open question, and Recursive's sources don't settle it. Its stated goal is recursively self-improving superintelligence, starting with model development. That's the kind of system a bill from Bernie Sanders and Greg Casar would ban (our story).

Recursive says its next work includes more capable local models and the infrastructure to run them. Flint also works with llama.cpp, Ollama, vLLM and SGLang, and the lab is releasing rec, its own inference engine for M4 Pro and M5 chips.