unsloth/Qwen3.8-27B
Qwen3.8-27B is a 27 billion parameter causal language model with a native vision encoder, developed by the Qwen Team. Built on the Qwen3.5 architecture, it delivers substantial improvements across coding, professional work, research, and long-horizon agentic tasks. This model features flexible thinking control and robust autonomous planning, making it highly reliable for complex, multi-step task completion. It also supports native image and video understanding, with a context length of up to 262,144 tokens.
Loading preview...
Qwen3.8-27B: Advanced Agentic and Multimodal Capabilities
Qwen3.8-27B, developed by the Qwen Team, is the latest and most capable generation in the Qwen open-model family, building upon the Qwen3.5 architecture. This 27 billion parameter model is designed for complex, multi-step tasks, offering significant advancements in agentic capabilities and multimodal understanding.
Key Capabilities and Enhancements
- Comprehensive Core Improvements: Enhanced performance across coding, professional work, research, and long-horizon agentic tasks.
- Robust Agent Execution: Features stronger autonomous planning and improved handling of environmental feedback, leading to more reliable end-to-end task completion.
- Flexible Thinking Control: Includes a default thinking mode with tunable reasoning depth via
reasoning_effortand retained historical reasoning context throughpreserve_thinking. - Native Vision-Language Understanding: Provides integrated support for understanding images and videos, ranging from STEM diagrams and documents to hour-scale video content.
- Extended Context Length: Natively supports a context length of 262,144 tokens, extensible up to 1,000,000 tokens for ultra-long texts using techniques like RoPE scaling.
Use Cases and Best Practices
Qwen3.8-27B is particularly well-suited for applications requiring advanced reasoning, complex task execution, and multimodal input processing. For optimal performance, specific sampling parameters are recommended for 'Thinking Mode' and 'Instruct Mode'. The model also benefits from adequate output length allocation for agentic tasks and specific configurations for long video understanding, such as setting longest_edge in the video preprocessor for higher frame-rate sampling.