Ollama Ships Flint: A 40B MoE Coding Model, macOS-Only, via MLX
Recursive's open-weight coding model lands in Ollama's MLX backend. One command to run it, one platform it works on, and a quantization path you can't shortcut.
What happened
Ollama v0.40.8 adds support for Flint, a 40B-parameter mixture-of-experts coding model from Recursive, through its MLX backend. One source tags the release v0.40.8-rc0, the other v0.40.8; the final tag is what ships. The command is ollama run flint, and it runs on macOS only for now.
Context
Recursive published Flint as an open-weight coding model: 40B total parameters, roughly 2.5B active per token, 262K-token context, reasoning and multi-turn tool use, Apache 2.0. Ollama has been the de-facto local-model runner for people who want ollama run <model> to just work, and its MLX backend has been the Apple-Silicon path for the last several releases. Adding a new MoE architecture to that backend is not a trivial patch. Flint required new attention primitives, a non-standard quantization guard, and a model-import pipeline that no other Ollama-supported model touches.
How it works
The integration is MLX-specific. The shared MLX attention layer gained two optional features: learned sinks and device-side causal masks. Existing models that call MLX attention keep their old no-sink, host-mask path; Flint opts into the new behaviour. During safetensors import, Ollama assembles ModelOpt expert banks for Flint, a step that does not exist in the import paths for other models. A guard rejects any attempt to create an NVFP4 tensor directly for Flint before conversion, forcing the safetensors-to-bank route. The model carries 40B total parameters with approximately 2.5B active per forward pass, so the full weight set must reside in unified memory even though only a fraction is touched per token. Other models retain their existing import and quantization paths; nothing about their pipeline changes.
Our read
The number everyone will quote is 74.2% on SWE-bench Verified. The number that matters more for a small studio is 54.9% on Terminal-Bench 2.1. SWE-bench Verified is a curated set of well-scoped GitHub issues; Terminal-Bench is closer to "sit in a shell and make the thing work." Fifty-five percent means roughly one in two terminal tasks succeeds on the first attempt. Useful assistant, not autonomous agent.
The release page also does not say what the NVFP4 guard costs in practice. If the only supported quantization path is the ModelOpt expert-bank assembly during import, the "just pull and run" simplicity that makes Ollama attractive adds a step you cannot skip. You cannot grab a pre-quantized NVFP4 file and drop it in.
And the macOS-only restriction is not a packaging choice. The attention path Flint requires, learned sinks and device-side causal masks, was written into the shared MLX layer. It is not a layer that exists in Ollama's CUDA path. If your build server is a Linux box, this release does not reach you, and "a subsequent release" with no date is the only commitment.
What this changes
If your dev machine is a Mac with 64 GB of unified memory, ollama run flint is available today and you can point it at a ComfyUI repo to trace a bug, scaffold a custom node, or write a test. If the box is under 32 GB, the 40B weight footprint will thrash and latency will hurt. If your engineering runs on Linux or NVIDIA hardware, nothing changes this release; Ollama defers other-GPU support to a subsequent release with no date. The video-generation stack is untouched. This is a developer-productivity tool, not a rendering one.
License
Flint is Apache 2.0-licensed. Commercial use is permitted without restriction; there is no revenue ceiling, no user-count cap, and no attribution requirement beyond what Apache 2.0 itself specifies. Code Flint writes for you carries no licence encumbrance.
Key takeaways
- Flint is a 40B MoE coding model from Recursive, roughly 2.5B active per token, 262K context, now runnable locally via
ollama run flinton macOS. - The MLX integration required new attention primitives (learned sinks, device causal masks) and a ModelOpt expert-bank import step that other Ollama models do not use.
- Direct NVFP4 creation is rejected for Flint; the safetensors-to-expert-bank path is the only supported quantization route.
- macOS-only at launch. Linux and NVIDIA GPU support is deferred with no published date.
- Reported scores: 74.2% SWE-bench Verified, 54.9% Terminal-Bench 2.1. Training-data provenance is not stated in any source.
Sources
- v0.40.8-rc0: mlx: add Flint model support (#19181) — tier 2
- Flint — Ollama v0.40.8 — tier 2
How this post was made
Drafted from clustered primary sources by the models below, then read, edited and approved by a human before it was published. The sources are listed in full at the end of the article.
- Drafted
- Independent sources
- 2
- cluster pair
- clef:27b
- cluster label
- gemma4:12b
- radar brief
- gemma4:12b
- research brief
- qwen3.8:27b
- draft article
- qwen3.8:27b
- short script
- qwen3.8:27b
- seo pack
- gemma4:12b
- Run
- editorial-20261022T205620Z