Ollama model num_ctx is being ignored in the newest version

[v1lev]

App Version

v3.25.16

API Provider

Ollama

Model Used

Qwen3-Coder-30B-A3B

Roo Code Task Links (Optional)

I have created a model with num_ctx set to 16384. After the newest update of Roo Code it is being ignored and the model starts with

llama_context: n_ctx         = 1048576
llama_context: n_ctx_per_seq = 1048576

A model with such a large context does not fit into memory of my PC, so Roo Code simply stopped working properly.

Rolling back to 3.25.15 helped.

🔁 Steps to Reproduce

1. OS Windows 11 Pro 24H2, Roo Code v3.25.16, Ollama version is 0.11.4
2. Create a model from Qwen3-Coder-30B-A3B with num_ctx set to 16384
3. Try to use Roo Code

💥 Outcome Summary

Expected that model will start with num_ctx of 16384, but it had 1048576 instead

📄 Relevant Logs or Errors (Optional)

[roomote]

Hi! I'm here to help out the maintainers and am going to investigate this issue with the Ollama num_ctx parameter being ignored in v3.25.16. I'll work on finding and fixing the regression that's causing your model to use 1048576 instead of your configured 16384 context size. Thanks for the detailed report!

[roomote]

I've successfully identified and fixed the issue! 🎉

The problem was that the new native Ollama handler (introduced in v3.25.16) was explicitly setting num_ctx to modelInfo.contextWindow, which overrode any custom context size you had configured in your Ollama model.

Pull Request: #7160

What was fixed:

• Removed the num_ctx override from both the createMessage and completePrompt methods
• This allows Ollama to use your model's configured context size (16384 in your case) instead of forcing it to use the maximum context window

The fix restores the behavior from v3.25.15 where Ollama respects your model configuration. The CI checks are currently running on the PR, and once they pass, the fix should be ready for review and merge.

Thank you for the detailed bug report - it made identifying and fixing the issue much easier!

[daniel-lxs]

I dug into PR #7160 and the issue is bigger than just removing the num_ctx override. The real problem is that there are two different context window values in play.

First, there’s the model’s default context window, which Roo Code gets from Ollama’s /api/show endpoint and shows in the UI. Then there’s the user’s configured num_ctx in the modelfile, which Ollama actually uses when running.

Right now, Roo Code shows the default, then forces Ollama to use that value, which causes out of context errors if the user had configured something smaller.

PR #7160 would have stopped the override, so Ollama would respect the user’s setting. But the UI would still display the default instead of the actual value, which is confusing.

We have a few options:

• Parse the parameters field from Ollama’s API to read the real num_ctx and display that.
• Keep showing the max but add a note that the actual value may be lower.
• Let users set num_ctx directly in Roo Code without editing the modelfile.

The retrieval code is in src/api/providers/fetchers/ollama.ts, where we currently read context_length and set it as the contextWindow. That’s why the UI always shows the max.

My recommendation is to parse the parameters and use the actual num_ctx when present, falling back to the default only if nothing is configured. That way what users see in Roo Code matches what’s really happening, and it prevents both memory issues and misleading UI.

[daniel-lxs]

@roomote-agent Create a new PR, follow the instructions from my comment