What is the setup for Ollama + GPT-OSS/Qwen3?

[TomLucidor]

Can't find the way to use small models to handle this issue https://github.com/vybestack/llxprt-code/blob/main/docs/settings-and-profiles.md
P.S. What about turning off thinking in favor of more experimentation?

[e2720pjk]

Just to clarify — are you working on the provider setup, or something else like model parameters?
If it's about the provider you can refer to this section
https://github.com/vybestack/llxprt-code?tab=readme-ov-file#using-local-models
Ollama is compatible with the OpenAI spec,so you can configure it the same way as other local models.

[TomLucidor]

Some models have reasoning vs non-reasoning mode, so I am thinking about that as well rather than just simple-set Ollama

[e2720pjk]

Did a quick check—this part is supported in gemini-cli, but the logic is hardcoded

llxprt-code/packages/core/src/core/client.ts

Lines 62 to 65 in b697609

	function isThinkingSupported(model: string) {
	if (model.startsWith('gemini-2.5')) return true;
	return false;
	}

and the model parameter only applies to Gemini model.

llxprt-code/packages/core/src/core/client.ts

Lines 583 to 586 in b697609

	thinkingConfig: {
	thinkingBudget: -1,
	includeThoughts: true,
	},

The tricky part is that each model has its own way of toggling "thinking" mode.
Since the parameter varies across models, trying to enumerate them all feels a bit pointless—unless there's a future API that standardizes the toggle, direct configuration isn’t really feasible.

I haven’t tested it myself, but if you want to configure a specific model, you might try using /set modelparam to see if it works.
e.g.,

/set modelparam enable_thinking true

for Qwen3-32B.
Ref:https://huggingface.co/Qwen/Qwen3-32B#switching-between-thinking-and-non-thinking-mode

[acoliver]

I also think there is another part where thinking tokens are getting filtered out and neither gettinf displayed or added to the context. We need: settings for enabling disabling (prob just /set modelparam as identified), settings for displaying/not (config/settings), settings for adding to context/not (prob /set) and display gglyphs.

I think:

1. Reasoning/thinking can be powerful
2. Its more powerful in some models when added to the context (sonnet/Opus have specific requirements about when/how)
3. It can be overwhelming for some models when displayed to users (deepseek especially)

[acoliver]

oh note we can do custom headers (that's how you turn it on in qwen)

[e2720pjk]

Follow-up on reasoning model support is happening in #398