Ollama context window triggers hallucinations

The reported "hallucinations" are likely a silent UX failure in Ollama's default configuration rather than a fundamental model flaw.

• –Silent truncation occurs when local files exceed the default `num_ctx` buffer, causing the model to lose the actual user instructions and "fill in the blanks."
• –Qwen3:4B is a robust model, but local inference performance is often bottlenecked by conservative configuration choices intended to preserve system RAM.
• –Users can resolve the issue by manually setting `PARAMETER num_ctx` in a Modelfile to 32k or higher, provided their hardware can support the memory overhead.
• –This highlights a critical need for local LLM runners to provide explicit warnings or UI indicators when input context is truncated.