ollama - 💡(How to fix) Fix MLX/NVFP4 models ignore num_ctx and always load with 256K context [2 comments, 2 participants]

What is the issue?

I found another issue that seems separate from the Codex App catalog metadata issue.

Environment:

• macOS
• Mac Studio M2 Ultra, 192GB unified memory
• Ollama v0.24.0
• Model: qwen3.6:27b-coding-nvfp4
• Custom model: qwen36-coding-fast

Modelfile:

FROM qwen3.6:27b-coding-nvfp4

PARAMETER num_ctx 16384 PARAMETER num_predict 512 PARAMETER temperature 0.1 PARAMETER top_p 0.9

ollama show qwen36-coding-fast --modelfile shows that num_ctx is set to 16384.

I also tried explicitly passing num_ctx at runtime:

curl -s http://localhost:11434/api/generate
-d '{"model":"qwen36-coding-fast","prompt":"","keep_alive":"30m","options":{"num_ctx":16384}}' >/dev/null

Before testing, I stopped the model with: ollama stop qwen36-coding-fast:latest

However, ollama ps still shows: qwen36-coding-fast:latest ... CONTEXT 262144

For comparison, the same approach works correctly with gpt-oss:20b: ollama ps shows the requested context size. So this appears to be specific to MLX/NVFP4 models, or at least qwen3.6:27b-coding-nvfp4.

Expected: The model should load with CONTEXT 16384.

Actual: The model loads with CONTEXT 262144 regardless of Modelfile num_ctx or runtime options.num_ctx.

This makes local Codex App usage much slower because the model is loaded with the full 256K context.

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response