Why It Happens
Ollama's /v1 OpenAI-compatible layer only forwards standard OpenAI parameters. Native Ollama options like num_ctx need to be passed inside an options: {} object in the raw request body. Most OpenAI-compatible clients don't do this, so the parameter is quietly dropped and Ollama falls back to the model's default context size.
The Fix
Bake num_ctx directly into the model via a custom Modelfile. This sidesteps the client entirely.
1. Export the existing Modelfile:
ollama show qwen3.6:27b --modelfile > qwen_custom.Modelfile
2. Add or update the num_ctx parameter in the file:
PARAMETER num_ctx 262144
3. Create a new model from it:
ollama create qwen3.6-256k -f qwen_custom.Modelfile
4. Update your client config to point to the new model name, then verify:
ollama ps
The context size shown in ollama ps should now reflect 262144 instead of 32768.