Qwen3.6-27B - Alibaba Tongyi's Open-Source Dense Multimodal Model

What is Qwen3.6-27B

Qwen3.6-27B is a 27-billion-parameter dense multimodal large model open-sourced by Alibaba's Tongyi Qianwen team, and is the only dense-architecture version in the Qwen3.6 series. The model focuses on agentic coding and multimodal reasoning, comprehensively surpassing the previous-generation 397-billion-parameter MoE flagship on coding benchmarks such as SWE-bench and Terminal-Bench Qwen3.5-397B-A17B, supports mixed input of images, video, and text, and can be deployed without MoE routing, greatly lowering the barrier to use for developers.

Main Features of Qwen3.6-27B

• Agentic Coding: Supports repository-level code reasoning, front-end workflows, and complex coding tasks, with integration into Claude Code、Claude Code and other mainstream coding assistants.
• Multimodal Understanding: Natively supports mixed input of images, video, and text, covering visual reasoning, in-depth document understanding, and interactive visual question answering.
• Dual-Mode Reasoning: Supports flexible switching between Thinking mode and Non-thinking mode to adapt to different task requirements.
• Chain-of-Thought Retention: Provides the preserve_thinking feature, which can retain the complete reasoning content of previous turns in messages, facilitating iteration on agentic tasks.
• Ultra-Long Context: Natively supports 262,144 tokens, expandable to 1,010,000 tokens, meeting long-document processing needs.

Technical Principles of Qwen3.6-27B

• Dense Architecture: Adopts a fully activated dense architecture with 27 billion parameters. Unlike MoE routing mechanisms, it requires no sparse activation during inference, making deployment more direct and simple.
• Multimodal Fusion: As a causal language model, it integrates a Vision Encoder to achieve native unified understanding of images, video, and text.
• Ultra-Long Context Reasoning: Supports a 200K context window in agentic coding evaluations, meeting repository-level code understanding needs.
• Hybrid Attention Layout: Adopts a hidden-layer design combining Gated DeltaNet and Gated Attention, integrating the advantages of linear attention and standard attention.

How to Use Qwen3.6-27B

• Online Experience: Visit Qwen Studio to interact directly without local deployment.
• Local Deployment
  • Download model weights from Hugging Face (Qwen/Qwen3.6-27B) or ModelScope.
  • Use mainstream frameworks such as vLLM, SGLang, and KTransformers to load and run it.
  • Switch to a domestic download source by setting an environment variable (such as VLLM_USE_MODELSCOPE=true).
• API Calls：Alibaba Cloud Model Studioplatform will soon launch the Qwen3.6-27B API service, supporting the preserve_thinking parameter.
• Coding Assistant Integration: Connect the model to OpenClaw, Claude Code, or Qwen Code to enable context-aware coding assistance.

Key Information and Usage Requirements for Qwen3.6-27B

• Release Date: April 22, 2026
• Model Specifications: 27 billion parameters, dense architecture (Dense)
• Open-Source License：Apache 2.0
• Context Length: API configuration supports 131072 tokens, and evaluation scenarios support a 200K context window
• Hardware Requirements: Deployable locally on a single or dual high-end GPU
• Multimodal support: Mixed input of images, video, and text
• API service: Coming soon to Alibaba Cloud Model Studio, with access supported in the Beijing, Singapore, and US regions
• Coding assistant compatibility: Supports OpenClaw (Node.js 22+), Claude Code, and Qwen Code (Node.js 20+)

Core advantages of Qwen3.6-27B

• Punching above its weight: The 27-billion-parameter dense model comprehensively surpasses the previous-generation flagship, a 397-billion-parameter MoE 15 times its size, on coding benchmarks.
• Deployment-friendly: The dense architecture requires no MoE routing, offering higher inference efficiency and significantly lowering the barrier to local deployment.
• Flagship-level coding capability: SWE-bench Verified 77.2, Terminal-Bench 2.0 59.3, and SkillsBench 48.2, all substantially ahead of the previous generation and models of the same scale.
• Aligned multimodal capability: Vision-language processing capability is consistent with Qwen3.6-35B-A3B, supporting both thinking and non-thinking modes.

Project addresses for Qwen3.6-27B

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

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

Application scenarios for Qwen3.6-27B

• AI-assisted programming: Provides developers with repository-level code understanding, bug fixing, and front-end development support to improve coding efficiency.
• Agent development: Serves as the core brain of an Agent, executing complex multi-step tasks and tool calls to build automated workflows.
• Intelligent document processing: Based on ultra-long context capability, performs in-depth document analysis, long report summarization, and cross-page information extraction.
• Visual question answering system: Handles mixed text-and-image content, enabling chart understanding, video analysis, and interactive visual dialogue.
• Enterprise local deployment: In data-privacy-sensitive scenarios, enables secure and controllable private AI services through open-source weights.