Most strong coding models live in someone else’s data center, and you rent them by the token. Recursive, a startup led by former Salesforce chief scientist Richard Socher, just released a model that goes the other way. Flint is a free, open-weight coding model built to run on a laptop, and the company says most of the work of building it was done by its own automated research system.
Both halves of that pitch are interesting, and both need some care. The laptop part is mostly a matter of arithmetic you can check. The “built by AI” part is a company claim that nobody outside Recursive has verified yet.
The short version
- Flint is a mixture-of-experts model with about 39.8B total parameters and only 2.5B active per token
- It ships under Apache 2.0 on Hugging Face, and quantized builds shrink to 11.4GB
- Recursive reports 89.7% on SWE-bench Pro v2 and about 131 tokens per second on an M4 Pro MacBook Pro
- The company says its Eureka Machine automated research system did most of the development work
- The catch: the benchmarks and the origin story are self-reported
What Flint actually is
Recursive released Flint on October 21, 2026, as its first model. It is a coding model first and a general chatbot second, aimed at developers who want an assistant that works offline, keeps source code on the machine and costs nothing per request.
| Spec | Flint |
|---|---|
| Architecture | Mixture of experts, 256 routed experts, 6 chosen per token |
| Parameters | 39.8B total, about 2.5B active per token |
| Context window | 262K tokens |
| License | Apache 2.0, weights on Hugging Face |
| Smallest quantized build | About 11.4GB |
| Reported speed | About 131 tokens/sec on M4 Pro MacBook Pro, 150+ on Nvidia RTX Spark |
The trick is the same one behind a lot of recent open models. A mixture-of-experts network stores a huge pool of specialist parameters but wakes only a small slice for each token. Flint holds nearly 40 billion parameters yet does the work of a 2.5 billion parameter model on every step, which is why it can be fast on consumer hardware.
Why a 40B model can fit on a laptop
Speed depends on active parameters, but memory depends on total parameters. That is the part people forget. All 39.8 billion weights still have to sit in memory even if only a few are used at once. At 16-bit precision that works out to roughly 80GB, which is data-center territory. Quantization is what changes the picture.
Recursive also quantized the model’s output layer to 4 bits, which goes against common practice because that layer is sensitive to precision loss. Its numbers suggest the gamble paid off: the NVFP4 build scores 88.5% on SWE-bench Pro v2 against 89.7% for the full-precision model, a drop of about a point.
The benchmark claims, read carefully
Recursive says Flint scores 89.7% on SWE-bench Pro v2, which it describes as behind only Alibaba’s Qwen 3.8 among the models it compared, and ahead of other coding models you can run locally. It also reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1.
Two things stand out. First, the gap between benchmarks is large: a model can look excellent on one test and merely decent on another. Second, Ollama’s listing for Flint explicitly notes that these are publisher-reported results and that evaluation setups can shift scores. Treat 89.7% as a starting hypothesis. If you plan to build a workflow around Flint, run it on your own repository and your own tasks first.
The “built by AI” claim
The more provocative part of the launch is how Flint was made. Recursive says its Eureka Machine, an automated research system, helped with nearly every step of development. It proposes ideas, runs experiments, reads the results and decides what to try next. Reported contributions include a low-bit training approach that cut the model’s size by 71% with little quality loss, and the company says the whole project took under four months.
How to read the claim
- Human researchers still supervised the process and picked the promising results
- “Largely built by AI” has no standard definition, so the share of human versus automated work is unknown
- Recursive frames Flint as an early test of recursive self-improvement, where AI helps build better AI
- Socher has said the long-term goal is full recursive self-improvement, with humans able to stop it
Automated research is a real and growing trend, and a 71% size reduction would be a meaningful result if it holds up. But “AI did it” is also good marketing, and the honest version of the story is probably a mix: automated experiments that explored a large search space, with people steering and filtering. Until Recursive publishes details, or outside researchers reproduce the training recipe, that is as far as the evidence goes.
Where Flint fits in the open model race
Flint lands in the middle of a week full of open-weight releases. Nvidia just argued that the next wave of agents will live on hardware you own, and we looked at how Nvidia squeezed a 74B model into 24GB to make that case. At the other extreme, Meta’s 816B open model needs a terabyte of RAM to run on a Mac. Flint sits between them: small enough for a normal developer machine, specialized enough to compete on coding tasks.
The hardware story matters too. Recursive’s 150-plus tokens per second figure comes from Nvidia’s RTX Spark platform, the same family inside the $2,600 Surface Laptop Ultra. Models like Flint are the reason machines like that exist.
Should you try it?
- Yes, if you want a private, offline coding assistant, have a laptop or desktop with 16GB or more of memory for a small quantized build, and are comfortable testing it yourself.
- Wait, if you need a drop-in replacement for a top cloud model today. Independent evaluations are only starting to appear, and the strongest claims come from the maker.
- Watch, if you care about research automation. Whether the Eureka Machine’s methods can be reproduced is the bigger story over the next few months.
A quick way to test it locally is through Ollama, which already lists Flint builds. Pull the smallest quantized version first and compare its output with your current assistant on a few real bugs from your own codebase:
ollama pull flint
ollama run flint "Find the bug in this function and explain the fix"The bottom line
Flint is a useful release even if every claim is only half true. A permissively licensed coding model that fits on a laptop lowers the cost of trying agentic coding and keeps your code on your own machine. The self-improvement narrative is the part to hold at arm’s length for now. Download the weights, run your own tests, and let the results, not the launch post, decide how seriously to take it.
Sources and further reading
- TechCrunch: Recursive launches Flint, a laptop coding model it says was largely built by AI
- MarkTechPost: Recursive releases Flint, an open-weight 40B MoE local coding model
- Ollama: Flint model library
- AI Weekly: Recursive opens Flint under Apache 2.0
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
