GitHub Repo
Ollama Preview Adds Flint Coding Agent Model, First Release Supports macOS Only
On October 21, Ollama released v0.40.8-rc0, adding local inference support for Recursive Flint through MLX. The open-weight model uses a sparse expert architecture, but the first release candidate supports only macOS; other GPU platforms are expected in later releases.
Ollama released v0.40.8-rc0 on October 21, adding Recursive’s coding agent model, Flint. The release notes list ollama run flint as the launch command and explicitly limit Flint support in rc0 to macOS, with support for other GPUs expected in later release candidates. For developers looking to run a coding agent locally, this offers a new deployment option, but they should check platform and version compatibility. Ollama release notes
Flint has 39.8 billion parameters, with about 2.5 billion activated per token. It supports text input and output, reasoning, and multi-turn tool calls, and its weights are licensed under Apache 2.0. The model configuration provides a context window of 262,144 tokens, shared by input and generated content. Sparse activation reduces computation per step, but it does not mean the full model’s memory requirements are equivalent to those of a 2.5-billion-parameter model. Official model card
The integration also involves tool prompting and attention computation. Ollama’s tag documentation says the implementation preserves the raw tool JSON alongside decoded API values, allowing Flint’s prompts to retain numeric types and schema order. MLX attention also adds optional learnable attention sinks and device-side causal masking. The import process assembles ModelOpt expert weights and rejects attempts to create a Flint NVFP4 model directly before conversion. Release technical notes
There was also a server-side deployment update that day. vLLM offers a dedicated Flint image and says its quantization kernel is optimized for a single Blackwell GPU; other configurations fall back to the standard layer. Therefore, the long-context decoding speeds reported by the team cannot be directly applied to Ollama or typical desktop GPUs. vLLM technical article
Next, it will be worth tracking when support for other GPUs arrives, as well as tool-call accuracy across different precision levels and runtime environments. The model card requires use of the official chat template and retention of reasoning_content in conversation history; it also cautions that long reasoning may exhaust the output budget. Engineering teams should evaluate completion rates, latency, and memory use on their own repository tasks. Context-window capacity alone does not guarantee quality on long tasks. Model usage limitations