Recursive released Flint on October 21, 2026: an open-weight, decoder-only mixture-of-experts model built for coding agents. It has 39.8 billion total parameters and about 2.5 billion active per token. The weights are on Hugging Face under Apache 2.0, and the model was free on OpenRouter at launch.
I would put it in a cheap evaluation slot for agent loops. I would not trust the headline benchmark numbers yet. Here is what the source actually gives us.
The specs in one table
| Item | Reported value |
|---|---|
| Total parameters | 39.8 billion (named 40B-A2.5B) |
| Active per token | ~2.5B including output head |
| Active model body | ~2.2B |
| Layers | 49 (2 dense + 47 MoE) |
| Context window | 262,000 tokens |
| License | Apache 2.0 |
| Price | Free on OpenRouter at launch |
| Terminal-Bench 2.1 | 54.9% |
| SWE-bench Pro v2 | 89.7% |
| SWE-bench Verified | 74.2% |
| Release date | October 21, 2026 |
The source credits MarkTechPost for the architecture and benchmark figures. Cryptopolitan is credited for the Eureka Machine claim.
What 2.5B active actually buys you
The first two layers are dense. The other 47 are mixture-of-experts layers that route each token to a small subset of experts. Inference compute therefore tracks a ~2.5B model rather than a 40B one. About 2.2B of the active parameters sit in the body. The rest is the output head, whose cost is largely fixed regardless of routing.
The catch is one the source only implies. Active parameters set per-token compute, not storage. All 39.8B parameters still have to live somewhere. The source says the design target is a single workstation or high-end laptop, but it gives no memory footprint, no quantization guidance and no official throughput. "Runs on a laptop" is the design intent, not a measured result.
If you have already read about on-device MoE models, this is the same trade. I covered a comparable one in the LFM2.5-24B-A2B post. For the heavyweight end of open coding MoEs, see Beam's 23B-active release.
Read the benchmark numbers carefully
All three scores are self-reported by Recursive. The source says independent reproduction was not available at publication, less than 24 hours after release.
The pattern in the scores is odd. SWE-bench Pro v2 is described as the harder, more realistic benchmark, yet Flint scores 89.7% on it. On SWE-bench Verified, the older and more established benchmark, it scores 74.2%. The source itself calls the Pro v2 number the standout and says a 2.5B-active model has "no obvious business" reaching that range if it holds up. I would treat the 74.2% Verified figure as the more conservative one. I would not plan around 89.7% until someone else reproduces it.
Two other claims are explicitly unverified in the source:
- 125+ tokens per second on an Apple M4 Pro MacBook Pro. This comes from social media or third-party reports, not Recursive or the model card.
- Compatibility with llama.cpp, Ollama, LM Studio, vLLM, Apple Silicon and RTX systems. None of this was established in the official sources the article reviewed.
Apache 2.0 weights make community support likely. Likely is not the same as working today.
The Eureka Machine claim
Reports say Recursive's automated research system, the Eureka Machine, helped develop Flint. Recursive has not published a breakdown of what it touched in design, training or evaluation. Treat it as the company's framing, not a finding. It has no bearing on how you evaluate the model.
How I would test it
The 262K context is the practical draw for agent work. It leaves room for several files, prior turns, tool outputs and test results in one window. Whether the model uses that context well is a separate question that only your own repo will answer.
- Point your existing agent harness at the free OpenRouter endpoint. Free access exists to get usage data, so expect it to be a launch-period offer rather than a lasting price.
- Run your own task set: bug fixes, feature implementation, test writing. These are the four jobs Recursive designed Flint around.
- Watch for loops and repeated failed fixes over long contexts, not just pass rate.
- Only then pull the weights locally and measure memory and tokens per second on your hardware.
The source reports that roughly one in four SWE-bench Verified tasks fails even at the stated 74.2%. For high-volume loops that call a model dozens or hundreds of times per task, a free Apache 2.0 model is worth benchmarking against whatever you pay for now. Do it as a cost-routing candidate for lower-stakes steps rather than a drop-in replacement.
Sources: