Tool Suggestions
Flint: The Open Coding Model That Runs at 125 Tokens per Second on a MacBook Pro
Recursive released Flint on October 21: a coding model built for agents that runs on the developer's own machine instead of a data center. The headline specification explains how — Flint holds 39.8 billion parameters but activates only about 2.5 billion of them for each token, so the compute per token stays low enough for a laptop. Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro, roughly twice the speed of comparably sized models, and the weights are public under an Apache 2.0 license with no revenue gate. It is also, the company says, a model largely built by AI — a claim worth separating from the model itself.
The specifications
Flint is a decoder-only mixture-of-experts transformer: 256 routed experts, six engaged per token, plus a shared expert, across 49 layers. Its documented context window is 262,144 tokens; Recursive says up to 512K is supported but recommends staying within 256K. The stated knowledge cutoff is June 2026. The weights are on Hugging Face as RecursiveSI/flint, and the model runs through the standard local stack — llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang — as well as Recursive's own rec engine, tuned for Apple Silicon.
The benchmarks — company-reported
Recursive reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1, and says Flint leads every mixture-of-experts model under 50 billion total parameters on agentic coding benchmarks while matching dense models of 25 to 40 billion parameters at more than five times their speed. Those figures come from Recursive's own evaluations, with setups that vary between labs, and no independent body had verified them at the time of writing. The speed numbers carry the same qualification: 125-plus tokens per second on an M4 Pro and 150-plus on a 64 GB NVIDIA RTX Spark are vendor measurements at short context lengths, though TechCrunch's report of Recursive's testing put the MacBook figure near 131 tokens per second against 58 to 69 for models such as Gemma 4 and Nemotron 3.5 Lightning on the same machine.
The limited hands-on evidence is encouraging. A Wired reporter who ran Flint on a Mac with 48 GB of memory gave it three jobs — a command-line to-do list, a web-search research tool and a simple explanatory website — and it completed all three quickly. Recursive itself adds a caveat in the other direction: Flint may need clearer, more detailed instructions than the large closed models developers are used to, and complex jobs should be split into smaller subtasks.
The hardware catch
"Runs locally" has a floor. Through the rec engine, Flint requires an M4 Pro, M5 Pro or M5 Max chip, at least 48 GB of unified memory and about 30 GB of free disk. The full BF16 checkpoint is roughly 80 GB to download; quantized builds are far smaller — the Ollama listing shows a 23 GB build — which is the realistic route for most laptops. Developers whose machines clear that bar get the actual prize: a capable coding agent that works offline, keeps code on the device, and sends no per-token bill to anyone.
Built by AI, in under four months
The second story is the process. Flint was produced by a small team of people and AI agents in under four months using Recursive's Eureka Machine, a system that proposes research ideas, writes the code to test them, checks the results and chooses the next experiment. Recursive says it contributed to nearly every part of the pipeline: it explored expert layouts before settling on the 40B/2.5B-active design, found a 4-bit quantization of the output layer that stays within 0.7 points of the full model, and wrote hardware-specific kernels for Apple Silicon and RTX Spark. "Nearly every aspect of the entire model pipeline was actually done and improved upon by AI," said CEO Richard Socher, Salesforce's former chief scientist and the founder of You.com. The company has raised $650 million at a $4.65 billion valuation, with backers including Google Ventures, Nvidia and AMD Ventures.
The measured assessment is that Flint the model and Flint the proof are different claims. The model is downloadable, benchmarked by its maker, and early-tested — judge it on your hardware, on your tasks. The claim that a self-improving system keeps getting better at building models is the company's long-term thesis, and Recursive has not shown outsiders an improvement curve for it yet. A good coding model built unusually fast is evidence of a strong pipeline; it is not yet evidence of recursion compounding.
The takeaway
For developers, Flint is immediately useful as an option: an openly licensed, genuinely fast local coding model for machines with 48 GB of memory or more, from a company with every incentive to keep improving it. For everyone else, it is a marker — the first public product of an AI-assisted model pipeline that a serious lab is willing to ship. Both are worth having, and neither needs the other to be true.