qwen2.5 coder Ollama returning weird output

[Happy GrassBlock (ImGGAAVVIINN)]

[Nikhil (shadowfax92)]

That garbled output is almost always one specific thing with Ollama: the context window is too small. Ollama defaults to 4,096 tokens, and below ~15K the agent's context overflows and it spirals into weird, looping output instead of clean responses. Fix is to start Ollama with a bigger context:

OLLAMA_CONTEXT_LENGTH=20000 ollama serve

Set 15,000–20,000 and it should straighten right out (heads up — it uses more VRAM; ollama ps shows your allocation).

Second thing: qwen2.5-coder is tuned for code completion, not the agent's tool-calling / structured output, so even with enough context it can be hit or miss. If it's still rough after the bump, try a model we've had good luck with — qwen/qwen3-4b is the easy starting point. Full setup + recommended list: https://docs.browseros.com/features/local-models

Give the context bump a shot and let me know if the weird output clears up?