Recursive just shipped Flint, a 40-billion-parameter open-weight coding model that runs at 131 tokens per second on a MacBook Pro. Locally. Offline. Free. The company achieved that speed by activating only 2.5 billion parameters per token — and by using an automated AI research loop to design the model itself.
The Speed That Makes This Interesting
Speed is the whole argument for Flint. Gemma 4 31B and Nemotron 3.5 Lightning run at 58–69 tokens per second on the same M4 Pro hardware. Flint hits 131 — roughly twice as fast. That delta matters more for coding agents than it does for chat: an agent running dozens of tool calls in a loop burns tokens constantly, and latency compounds. At 131 tok/s, a local agent loop becomes usable. At 60, it becomes something you watch and wait for.
Recursive achieves that speed through a mixture-of-experts architecture pushed to an extreme: 39.8 billion total parameters, but only 2.5 billion activate per token. Each transformer layer contains 256 routed experts; only 6 get selected per token, plus one always-active shared expert. The model carries the knowledge of a 40B system but, during inference, computes like a 2.5B one. Custom GPU kernels — designed partly by Recursive’s own automated research system — handle the dispatch efficiently enough to hit those throughput numbers.
How AI Built This Model
The more interesting story is how Flint came to be. Recursive built it in under four months using something they call the Eureka Machine: an automated research loop where agents propose architectural changes, run experiments, evaluate results, and feed the findings into the next iteration. According to Recursive’s official announcement, the Eureka Machine drove decisions on architecture shape, quantization (it found a 4-bit recipe that shrank the model by 71% with minimal quality loss), and the custom kernels that make the speed possible.
Founder Richard Socher framed Flint as the first public evidence the Eureka Machine works at scale: “The bigger test is whether AI can build the next one.” If the loop holds, Flint is not the product — it is the proof of concept for the process. No detailed paper has been published yet, so the claims are unverified externally, but the benchmark results are real.
Benchmarks: Best-in-Class With Caveats
Flint leads the sub-50B mixture-of-experts cohort on agentic coding benchmarks: 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2, and 54.9% on Terminal-Bench 2.1. It beats Qwen 3.6 35B-A3B (its direct MoE competitor) on five of six benchmarks, including by 10 points on Terminal-Bench.
| Model | Active Params | Tok/s (M4 Pro) | SWE-bench Pro v2 |
|---|---|---|---|
| Recursive Flint | 2.5B | 131 | 89.7% |
| Qwen 3.6 35B-A3B | 3B | ~90 | lower |
| Gemma 4 31B | 31B (dense) | 58–69 | n/a |
The honest footnote: Qwen 3.8 27B, a dense model in a different category, beats Flint on four of six benchmarks — including Terminal-Bench 2.1 by 18 points. Recursive’s benchmark framing carefully avoids that comparison. If you are optimizing purely for coding quality and have no hardware constraint, Qwen 3.8 is still the better choice. Flint’s value proposition is speed plus privacy, not raw benchmark supremacy.
Running It
Flint is on Hugging Face under Apache 2.0. The easiest path is Ollama:
ollama run flint:40b-a2b-qat-int4That pulls a 23 GB model. At 4-bit quantization, you need 24–27 GB of RAM or unified memory. Recursive’s native inference engine — the one that produces the 131 tok/s figure — currently supports only M4 Pro, M5 Pro, and M5 Max Macs with 48 GB of unified memory. A 24 GB M4 Pro MacBook will run it via Ollama but not at peak speed. NVIDIA users need 26 GB of VRAM for 4-bit, 44 GB for 8-bit.
Once running, Flint exposes an OpenAI-compatible API. Point Cline, Claude Code, or OpenCode at http://localhost:11434 with model ID flint:40b-a2b-qat-int4. TechCrunch’s coverage has additional setup context.
The Bottom Line
Flint is the best local MoE coding model available today. It is not the best local coding model overall — Qwen 3.8 edges it on raw benchmarks — but it closes most of that gap in practice by being significantly faster at inference. For agentic workflows, offline environments, or teams that cannot send code to an external API, it is a serious option worth evaluating. And if the Eureka Machine works as advertised, Recursive’s next model will be better, built faster, and probably more benchmark-honest. The weights are free — there is no reason not to try it.
