Qwen3.6-Plus: agents got smarter, but benchmarks are still misleading 🤖

Alibaba released Qwen3.6-Plus. The team claims their model is very good at solving problems at the level of entire repositories. However, in their tables, they shyly compare themselves with GLM-5 and Opus 4.5, not 5.1 and 4.6.

If we take their own benchmark for working with repositories via Claude Code (NL2Repo) and put next to it the results of the recently released GLM-5.1, we see:
▫️ Claude Opus 4.6 — 47.9
▫️ GLM-5.1 — 45.3
▫️ Qwen3.6-Plus — 37.9

For serious work with legacy code and large refactorings, GLM-5.1 is much preferable.

👍 What's good: they added the preserve_thinking parameter to the API.

How it works for others now: in a classic agent loop, the model generates reasoning (its thoughts) at each step, which we either discard or feed back, burning a ton of tokens.
With preserve_thinking=True, the API itself saves the reasoning context from all previous dialog steps. This radically reduces costs and prevents agent "schizophrenia" on long multi-step tasks. By default, the parameter is off; it is passed via extra_body in the request. Keep this in mind if you are building your own agents.

They also released integration with OpenClaw, the Qwen Code CLI tool, and native support for the Anthropic API format. That is, you can directly plug their endpoint into Claude Code and work from the terminal. Plus a default context window of 1M tokens.

Overall, another solid multimodal model with cool features for API developers from the Chinese, but without the "wow" factor.

You can already try it in the chat. 💬