Qwen3.6-Plus is now officially available via Alibaba Cloud Model Studio, marking a major step forward in the development of native multimodal artificial intelligence agents. Built on feedback from the Qwen3.5 era, this release delivers sweeping improvements across agentic coding, multimodal vision, and general reasoning, all backed by a 1M token context window enabled by default.
What Makes Qwen3.6-Plus Different
The model is designed from the ground up for real-world developer workflows. It handles everything from frontend web development to complex, repository-level problem solving. The team describes the experience as “vibe coding”: a fluid, intuitive interaction between developer and model that feels less like issuing commands and more like collaborating with a capable engineer.
Four headline features define this release:
- 1M Context Window: Available by default via the API, supporting up to 991K input tokens.
- Agentic Coding: Dramatically improved performance on SWE-bench, Terminal-Bench, and front-end code generation tasks.
- Multimodal Vision: Sharper object recognition, OCR, document understanding, and spatial reasoning.
- Deep Thinking: A new
preserve_thinkingAPI parameter retains reasoning context across multi-turn agent tasks, reducing redundant computation.
Benchmark Performance for Qwen3.6-Plus
The numbers tell a clear story. On Terminal-Bench 2.0, the model scores 61.6, outperforming Claude Opus 4.5 (59.3), Kimi-K2.5 (50.8), and GLM5 (56.2). On QwenClawBench, an internal real-user-distribution coding agent benchmark, it leads with 57.2 against Claude Opus 4.5 at 52.3. For general agents, it tops TAU3-Bench at 70.7, edging out Claude Opus 4.5 at 70.2. On MCPMark, it scores 48.2, ahead of all listed competitors.
Vision performance is equally strong. On OmniDocBench 1.5, the model reaches 91.2, the highest among all compared models. It scores 97.6 on CountBench and 93.5 on RefCOCO (avg), both top results. On OSWorld-Verified, a visual agent benchmark, it scores 62.5, competitive with Claude Opus 4.5 at 66.3 and ahead of Kimi-K2.5 at 63.3.
Coding Tools and Integrations
The model works with a wide range of developer tools out of the box. Supported integrations include OpenClaw, Claude Code, Qwen Code, Kilo Code, Cline, and OpenCode. The API is compatible with both OpenAI and Anthropic protocol standards, so teams can drop it into existing pipelines with minimal friction.
Qwen Code users get 1,000 free calls per day after signing in via OAuth. OpenClaw users can self-host the open-source agent and point it at Model Studio with a simple JSON configuration. Claude Code users can redirect the Anthropic client to the Alibaba Cloud endpoint using environment variables, no code changes required.
Multimodal Vision and Agent Capabilities
Beyond AI coding, the model advances along three dimensions in multimodality: reasoning, real-world applicability, and agentic execution. It can analyze documents, parse charts, understand UI screenshots, and generate frontend code from design mockups. In GUI agent scenarios, it reads the current screen state, plans the next action, and executes it, moving closer to a fully autonomous software agent.
Video understanding is also improved. The model handles temporal reasoning across frames, supports lecture note generation from video, and can assist with video editing workflows. These capabilities position it well for enterprise use cases in retail intelligence, document processing, and automated software testing.
Pricing and Availability
The model is priced at $0.50 per million input tokens and $3.00 per million output tokens. Explicit cache creation costs $0.625 per million tokens, while cache reads drop to $0.05 per million. Web search tool calls are priced at $10 per 1,000 calls. The code interpreter and web extractor tools are currently free during a limited promotional period.
The Qwen team has confirmed that smaller open-source variants from the Qwen3.6 series are coming soon. For businesses and developers in the region looking to integrate cutting-edge AI into their stacks, the model is accessible now through the Singapore endpoint at Alibaba Cloud Model Studio.
