Recursive's new open weight Flint runs fast on a laptop, but it is a reasoning model: starve its output budget or strip its reasoning from history and it will look broken. Four settings fix it.
Recursive's Flint is the most interesting local coding model of the week: roughly 40 billion total parameters, about 2.5 billion active per token, Apache 2.0 weights, and more than 125 tokens per second on an M4 Pro MacBook Pro at short contexts, by Recursive's own measurements. The catch is that Flint is a reasoning model, and most local harnesses are configured for models that answer immediately. Run it with your old defaults and it will look slow, truncated and forgetful. Configure it properly and it behaves like the model in the benchmarks.
Start with the output budget. Flint's model card recommends at least 32,768 output tokens for hard problems, and that budget includes the reasoning it writes before it answers. The model card also warns that long reasoning can use up the output budget before an answer finishes. Give it a small cap on a real refactor and it can run out while still thinking, which looks like a broken model when it is really a starved one.
Next, keep its reasoning in the conversation. The model card tells you to preserve the returned reasoning_content in conversation history. Check that your client actually sends that field back on the next turn rather than discarding it as unfamiliar. In a multi turn agent loop, dropping it means Flint loses track of the plan it made one step earlier. Flint supports OpenAI compatible tool calls and multi turn tool use, so this matters most exactly where you want it most: agentic coding.
Then match the sampling Recursive evaluated with. Flint's results use temperature 1.0, top_p 0.95 and top_k 20. If you lower the temperature out of habit, you are running a configuration other than the one behind the published results.
Finally, respect the context guidance. The documented context configuration is 262,144 tokens, shared between prompt and output. Recursive says Flint can stretch to 512k but recommends staying within 256k for best results. Leave headroom for that 32,768 token output budget, and do not stuff a whole monorepo into the prompt just because the window allows it.
Try it today
- Set max output tokens to at least 32,768 for hard coding tasks, since the budget includes Flint's reasoning.
- Make sure your client sends the returned reasoning_content back in conversation history on every turn.
- Use temperature 1.0, top_p 0.95 and top_k 20, the settings Recursive's published results use.
- Keep prompt plus output within 256k tokens even though the documented window is 262,144, and reserve room for the output budget.
- Break large tasks into smaller subtasks; Recursive says Flint may need more explicit guidance than larger closed models.