Recursive's Flint Is a Free Apache 2.0 Coding Model for Laptops. Every Benchmark Is Recursive's, and "Built by AI" Comes With No Number.
Recursive released Flint on October 21: a 39.8B-total, roughly 2.5B-active mixture-of-experts coding model under Apache 2.0 on Hugging Face, with a 262K context and a June 2026 cutoff. The model card reports Terminal-Bench 2.1 at 54.9%, SWE-bench Pro v2 at 89.7%, and SWE-bench Verified at 74.2%, all measured by Recursive. The company says its Eureka Machine helped build nearly every part of the model, with no percentage attached. Artificial Analysis has not scored it. We read Recursive's post, the Hugging Face card, and the launch coverage.
By T.W. Ghost
Recursive, Richard Socher's lab, released Flint today, October 21, 2026: a mixture-of-experts coding model with 39.8 billion total parameters and about 2.5 billion active, on Hugging Face under Apache 2.0, pitched as a model that runs on a laptop. The blog quotes speeds above 125 tokens per second on an M4 Pro and above 150 on an RTX Spark; the Hugging Face card's deployment kit, by contrast, targets B300 and H200 servers running SGLang. Wired's hands-on used a 48 GB Mac.
The scores on the card are the kind that stop a reader: Terminal-Bench 2.1 at 54.9%, SWE-bench Pro v2 at 89.7%, SWE-bench Verified at 74.2%. Every one carries a dagger, and the dagger means Recursive measured it. Recursive also says its Eureka Machine, an AI research system, "helped with nearly every step" of building Flint, a claim the launch coverage led with and which comes with no percentage of the work. Artificial Analysis has no entry for Flint on any board. We read the launch post, the model card, and the TechCrunch coverage. Here is what held up.
What You Are Downloading
| Recursive Flint | |
|---|---|
| Released | October 21, 2026; RecursiveSI/flint on Hugging Face |
| License | Apache 2.0 |
| Architecture | Mixture of experts; 39.8B total; about 2.5B active (including the output head) |
| Context | 262K in the config |
| Knowledge cutoff | June 2026 |
| Download | About 80 GB in BF16 |
| Vendor laptop speeds | Above 125 tokens/s on M4 Pro; above 150 on RTX Spark (blog) |
| Reference deployment | SGLang on B300 and H200 (card) |
| AA entry | None found |
Apache 2.0 with no gate puts Flint in the same license tier as Reflection's Beam and above Meta's Muse Spark Open, whose license has a revenue clause. At 40B total it is also the only one of the three that a well-equipped laptop can hold.
The speed claims are vendor claims. The blog's laptop figures and the card's server kit describe different hardware; we print both and note that the only outside hands-on we have, Wired's, ran on a 48 GB Mac.
The Benchmarks, All of Them Recursive's
| Benchmark (Recursive's model card) | Flint |
|---|---|
| Terminal-Bench 2.1 | 54.9% |
| SWE-bench Pro v2 | 89.7% |
| SWE-bench Verified | 74.2% |
| Method | Agentic average pass@1 over at least 5 runs; mini-swe-agent harness |
Three honest observations:
- Every hard number is the vendor's. The dagger on the card marks Recursive-measured results. There is no AA score, no third-party Terminal-Bench run, no independent SWE-bench run.
- 89.7% on SWE-bench Pro v2 would be remarkable, which is why it needs a second source. Label it with the harness (mini-swe-agent) and the measurer (Recursive) until someone else reproduces it.
- Terminal-Bench 2.1 is the older, easier version, and Recursive's scoring note says Flint's web-retrieval solutions were counted as unsuccessful, which is conservative in Flint's disfavor. It is still version 2.1, not 4.0, and not comparable to the 4.0 figures on this site.
On "built by AI": the launch post says the Eureka Machine helped with nearly every part of building Flint. No share of the work, no description of what it did versus what people did. It is a qualitative claim and we treat it as one.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Oct 21; Apache 2.0; 39.8B total / about 2.5B active; 262K context; June 2026 cutoff; about 80 GB BF16 | Verified (Hugging Face card and repo) |
| Terminal-Bench 2.1 54.9%; SWE-bench Pro v2 89.7%; SWE-bench Verified 74.2% | Vendor (Recursive-measured, on the card) |
| Laptop speeds above 125 and 150 tokens/s | Vendor (blog); Wired ran it on a 48 GB Mac |
| Eureka Machine built nearly every part | Vendor, qualitative, no percentage |
| Independent score on any board | Unverified (none found) |
Who Should Use It
Download Flint if you want a permissively licensed coding model you can run locally at about 80 GB and evaluate on your own repository; the license and the size are the verified parts of the story.
Wait before ranking it against Beam or Muse Spark Open on capability; there is no independent measurement, and the scores on the card are the vendor's. Take the architecture numbers from the Hugging Face card, not the blog.
The frontier is elsewhere: on the open-weight boards Artificial Analysis publishes, the US entries this month are Muse Spark Open at 39 and Beam at 30, both well under the Chinese open models, and Flint is not on the board at all. Our model comparison has the current standings side by side.
Sources
- Recursive: introducing Recursive Flint
- Hugging Face: RecursiveSI/flint
- TechCrunch: Recursive launches Flint
*Not sure whether a local coding model is worth the disk space before anyone independent has scored it? Take the free 2-minute quiz and get matched. Then read about the other permissively licensed US open model this month, at twelve times the size: Reflection Beam.*