Ollama v0.30.9 adds Cohere2Moe, fixes Claude agent one-token bug

Ollama, the popular open-source local LLM runner with 174k GitHub stars, released v0.30.9. The update introduces support for the Cohere2Moe architecture, expanding the range of models users can run locally. More critically, it fixes a significant bug where launching Claude and other coding agents via Ollama would only output a single token — a blocking issue for agentic workflows that require multi-step responses. The LFM2 parser/render bug that prevented thinking tokens from being emitted has also been resolved. Additionally, Ollama now returns a clear error if a single message exceeds the current context window, preventing silent failures.

This release addresses pain points for developers integrating Ollama into coding agent pipelines, where tools like Claude need to generate coherent multi-turn outputs. By fixing the one-token output bug and improving error handling, Ollama v0.30.9 makes local AI agents more reliable for real-world use. The Cohere2Moe support also signals Ollama's commitment to supporting diverse model architectures beyond Llama variants. With its massive community (16.7k forks), Ollama continues to be a cornerstone for running AI models on consumer hardware.