Alibaba opens the next-generation flagship Qwen3.6-Max-Preview preview, focused on agent-based programming

ME News, April 20 (UTC+8): According to Beating Monitoring, the Alibaba Qianwen team released Qwen3.6-Max-Preview. It is positioned as an early preview of the next-generation flagship model, replacing the existing Qwen3.6-Plus. Users can chat and try it directly in Qwen Studio (chat.qwen.ai). They can then open API access via Alibaba Cloud Bailing using the model name qwen3.6-max-preview. The interface is compatible with OpenAI’s chat completions and responses formats, as well as Anthropic’s protocols.

This version is mainly aimed at agentic coding, enabling the model to write code, run it, view errors, call tools, and complete multi-step programming tasks like a programmer. Compared with the previous Qwen3.6-Plus, the official says the improvements are concentrated in programming: SkillsBench +9.9, SciCode +10.8, NL2Repo +5.0, and Terminal-Bench 2.0 +3.8 points. World knowledge and tool-calling formatting follow with improvements ranging from 2.3 to 5.3 points across the other three metrics. The official claims it achieved the highest scores on six programming benchmarks, including SWE-bench Pro, Terminal-Bench 2.0, and SciCode. Among them, QwenClawBench and QwenWebBench, judging by their names, appear to be Qianwen’s own internally built evaluation sets, which should be considered separately from public benchmarks.

On the API side, a new preserve_thinking option has also been added: when enabled, messages will retain the thinking content from the previous few rounds. The reasoning model normally returns only the “thinking of the current round” once per turn. In multi-turn agent conversations, the context lacks the thinking from earlier steps, so when the model replans it may head back down previously tried paths or forget what it has already tried. This switch is intended to address that gap. (Source: BlockBeats)