Alibaba releases Qwen3.6-Plus with 1M context and Code Arena #8

• Alibaba launched Qwen3.6-Plus as its new hosted flagship, with a 1 million token context window, stronger agentic coding, and upgraded multimodal reasoning launch thread.
• Qwen’s own evals put it at 61.6 on Terminal-Bench 2.0, 56.6 on SWE-bench Pro, 78.8 on SWE-bench Verified, and 57.2 on QwenClawBench, a very aggressive coding-agent jump over Qwen3.5-397B-A17B LM benchmark table.
• The vision side moved too: Qwen3.6-Plus posted 91.2 on OmniDocBench 1.5, 85.4 on RealWorldQA, and 87.8 on Video-MME with subtitles VLM benchmark table.
• Outside Alibaba’s own charts, Arena said the preview model opened at #8 overall on its Code Arena leaderboard, with Alibaba becoming the #2 lab on the React leaderboard Code Arena ranking.
• Access expanded fast on launch day: Qwen shipped on OpenRouter as a free model with no prompt retention during the promo period OpenRouter launch, and Alibaba also announced upcoming Fireworks support for inference and fine-tuning Fireworks partnership.

The official post says Qwen3.6-Plus adds a new preserve_thinking API option for multi-step agent runs. Alibaba’s model catalog lists the hosted model at 1M context with text, image, and video input support. You can also pull it through OpenRouter, check the Code Arena leaderboard, or use Vercel’s AI Gateway model ID alibaba/qwen3.6-plus.

Alibaba pitched this one around coding agents, not general chatbot polish.

The internal benchmark sheet splits the gains into a few distinct buckets:

• Code repair and repo work: 78.8 on SWE-bench Verified, 56.6 on SWE-bench Pro, 37.9 on NL2Repo.
• Terminal automation: 61.6 on Terminal-Bench 2.0.
• Real-world agent tasks: 58.7 on Claw-Eval pass^3, 57.2 on QwenClawBench.
• Web artifacts: 1501.7 Elo on QwenWebBench.
• Tool use and planning: 48.2 on MCPMark, 74.1 on MCP-Atlas, 41.5 on DeepPlanning.

The interesting bit in the official post is how Alibaba describes the change: deeper integration of reasoning, memory, and execution, plus explicit support for agent scaffolds like OpenClaw, Claude Code, Qwen Code, Kilo Code, Cline, and OpenCode.

The launch thread made the 1M window a headline feature, but the blog adds the more useful implementation detail: Qwen3.6-Plus exposes a preserve_thinking flag that keeps reasoning content from earlier turns in the message history for agentic tasks.

According to the official announcement, the flag is off by default. Alibaba says preserving prior thinking can improve decision consistency on long multi-step runs and can sometimes reduce token use by avoiding repeated reasoning.

Alibaba’s Model Studio catalog lists Qwen3.6-Plus as a hosted multimodal model with text, image, and video inputs, a 1,000,000 token context window, and minimum pricing of 2 yuan per million input tokens and 12 yuan per million output tokens across its listed deployment regions.

The second part of the release is that Qwen is trying to make the coding story visual.

Alibaba’s table shows three clusters of gains:

• Document and OCR work: 91.2 on OmniDocBench 1.5, 83.4 on CC-OCR.
• General image reasoning: 85.4 on RealWorldQA, 94.4 on AI2D_TEST.
• Video: 87.8 on Video-MME with subtitles, 84.2 without subtitles.

The blog post frames that as a path toward visual coding workflows, where a model can read screenshots, design comps, or UI state, then generate or modify code with tools. That is Christmas-come-early energy for people who care about front-end agents more than benchmark purity.

Alibaba’s benchmark deck is strong, but launch-day third-party positioning came from Arena.

Arena said Qwen 3.6 Plus Preview debuted at #8 overall on its Code Arena leaderboard. The linked leaderboard describes the ranking as votes on agentic coding tasks involving multi-step reasoning and tool use, and the screenshot in the tweet shows a preliminary score of 1454.

Alibaba amplified the result in its own follow-up posts Alibaba on Code Arena, but the more interesting signal is where the model landed: close enough to frontier proprietary coding models to show up on the same board immediately.

Qwen pushed distribution hard instead of keeping the model boxed inside Alibaba Cloud.

OpenRouter said Qwen 3.6 Plus was available for free at launch and that prompts and completions were not retained during that period OpenRouter launch. Vercel also published same-day support in AI Gateway, with the model ID alibaba/qwen3.6-plus. Alibaba separately said the model was live in Qoder Qoder support.

That rollout matters because the launch is really about agent surfaces. Qwen is easier to test when it appears inside existing coding tools instead of only through one vendor API.

Alibaba ended the launch day by teeing up another distribution channel.

The Fireworks announcement says Qwen3.6-Plus is coming to Fireworks AI for both inference and fine-tuning, not just hosted chat access Fireworks partnership. That adds a different deployment angle than the OpenRouter and Vercel integrations, which were about immediate routing and API access.

It is also one of the few concrete post-launch facts not already in the main thread: Alibaba is treating Qwen3.6-Plus as a model to be served, routed, and tuned across multiple stacks on day one, not just a benchmark trophy.