Qwen3.6-27B - Dense Multimodal Model Open Sourced by Alibaba Tongyi

What is Qwen3.6-27B?

Qwen3.6-27B is a dense multimodal model with 27 billion parameters, open-sourced by Alibaba's Tongyi Qianwen Team. It is the only dense architecture version in the Qwen3.6 series. The model focuses on agent programming and multimodal inference, and comprehensively surpasses the previous generation's 397 billion parameter MoE flagship Qwen3.5-397B-A17B on programming benchmarks such as SWE-bench and Terminal-Bench. It supports mixed input of images, videos, and text, and can be deployed without MoE routing, significantly reducing the barrier to entry for developers.

Main functions of Qwen3.6-27B

• Intelligent agent programmingSupports repository-level code reasoning, front-end workflows, and complex programming tasks, and integrates mainstream programming assistants such as Claude Code.
• Multimodal understandingIt natively supports mixed input of images, videos, and text, covering visual reasoning, deep document understanding, and interactive visual question answering.
• Dual-mode reasoningIt supports flexible switching between thinking and non-thinking modes to adapt to different task requirements.
• Mind chain retention:supply preserve_thinking This feature allows the retention of complete reasoning content from previous rounds within messages, facilitating task iteration for the intelligent agent.
• Extremely long contextIt natively supports 262,144 tokens, which can be expanded to 1,010,000 tokens, meeting the needs of long document processing.

Technical Principles of Qwen3.6-27B

• Dense architectureIt adopts a dense architecture with 27 billion parameters and full activation, which is different from the MoE routing mechanism. It does not require sparse activation during inference, making deployment more direct and simple.
• Multimodal fusionAs a causal language model integration with Vision Encoder, it enables native unified understanding of images, videos, and text.
• Extremely long contextual reasoningSupports 200K context windows in agent programming evaluation to meet the needs of repository-level code comprehension.
• Hybrid attention layoutIt adopts a hidden layer design that combines Gated DeltaNet and Gated Attention, integrating the advantages of linear attention and standard attention.

How to use Qwen3.6-27B

• Online experienceAccess Qwen Studio directly for interactive dialogue, without the need for local deployment.
• Local deployment
  • From Hugging Face (Qwen/Qwen3.6-27BAlternatively, you can download the model weights using ModelScope.
  • It can be loaded and run using mainstream frameworks such as vLLM, SGLang, and KTransformers.
  • By setting environment variables (such as VLLM_USE_MODELSCOPE=trueSwitch to a domestic download source.
• API callsAlibaba Cloud's Bailian platform will soon launch the Qwen3.6-27B API service, supporting... preserve_thinking parameter.
• Programming Assistant IntegrationConnect the model to OpenClaw, Claude Code, or Qwen Code to achieve context-aware coding assistance.

Key information and usage requirements of Qwen3.6-27B

• release dateApril 22, 2026
• Model Specifications27 billion parameters, dense architecture.
• Open source licenseApache 2.0
• Context lengthAPI configuration supports 131,072 tokens, and evaluation scenarios support 200K context windows.
• Hardware RequirementsHigh-end GPUs, whether single or dual-card, can be deployed locally.
• Multimodal supportMixed input of images, videos, and text
• API ServiceAlibaba Cloud's "Hundred Refinements" service is about to launch, supporting access from Beijing, Singapore, and the United States.
• Programming assistant compatibility: Support OpenClaw (Node.js 22+), Claude Code, Qwen Code (Node.js 20+)

Qwen3.6-27B's core advantages

• Winning big with small investmentThe 27 billion parameter dense model surpasses the previous flagship MoE, which had 397 billion parameters and was 15 times larger, across all programming benchmarks.
• Deployment friendlyDense architectures eliminate the need for MoE routing, resulting in higher inference efficiency and significantly lower barriers to local deployment.
• Flagship-level programming capabilitiesSWE-bench Verified 77.2 points, Terminal-Bench 2.0 59.3 points, SkillsBench 48.2 points, all significantly outperforming previous generations and models of the same scale.
• Multimodal capability alignmentThe visual language processing capabilities are consistent with Qwen3.6-35B-A3B, supporting both thinking and non-thinking modes.

Project address of Qwen3.6-27B

• Project official website: https://qwen.ai/blog?id=qwen3.6-27b
• HuggingFace model libraryhttps://huggingface.co/Qwen/Qwen3.6-27B

Comparison of Qwen3.6-27B with similar competing products

Application scenarios of Qwen3.6-27B

• AI-assisted programmingIt provides developers with repository-level code understanding, bug fixing, and front-end development support, improving coding efficiency.
• Intelligent agent developmentAs the core brain of the Agent, it executes complex, multi-step tasks and tool calls to build automated workflows.
• Intelligent document processingLeveraging its ability to provide extensive contextual information, the system enables in-depth document analysis, long report summaries, and cross-page information extraction.
• Visual question answering systemIt can handle mixed text and graphics content, enabling chart understanding, video analysis, and interactive visual dialogue.
• Enterprise on-premises deploymentIn data privacy-sensitive scenarios, secure and controllable private AI services can be achieved through open-source weights.