Life quality tip for everyone using Qwen3.6-35B-A3B

[HLGRL]

Hello everyone,

Qwen team introduced this feature in Qwen 3.6, this is quote from their huggingface card:

Preserve Thinking

By default, only the thinking blocks generated in handling the latest user message is retained, resulting in a pattern commonly as interleaved thinking. Qwen3.6 has been additionally trained to preserve and leverage thinking traces from historical messages. You can enable this behavior by setting the preserve_thinking option:

To avoid constantly repeating loops and failed tool calls for example in Claude Code, when using Qwen3.6-35B-A3B please do this step in oMLX.

In "Chat Template Kwargs" click "Add" and select "Custom" then:
in key write: preserve_thinking
in value write: True
click "Save".
After this you should have clear and smooth responses in Claude Code without any stuck reasoning loops or failed tool calls.

@jundot Can you somehow make it as a default option for every quant/variation of Qwen 3.6 in oMLX?
Without that "preserve_thinking" setting after some context it gets stuck in reasoning loops and it is not enabled by default in model card, and most people don't know about this, but that improve LLM output quality a lot!

Thanks

[deepsweet]

I believe it's a bit more complex than just adding the kwarg: #814

Wiring preserve_thinking=True server-side isn't enough though: external OpenAI/Anthropic-compatible clients receive clean content (thinking is split into reasoning_content by oMLX's reasoning parser) and echo back only content on the next turn — so there's nothing to preserve.

This PR wires both halves: the server auto-enables preserve_thinking when the template supports it, and the API layer reconstructs blocks from client-echoed reasoning_content / Anthropic thinking blocks before templating.

Waiting for the PR to be merged 👀