Qwen3.8 27B (Non-reasoning)

Qwen3.8 27B is a dense, open-weight multimodal model developed by Alibaba and released in August 2026. As a core member of the Qwen3.8 series, it is designed to bridge the gap between compact local models and frontier-class systems, specializing in coding, professional research, and long-horizon agentic tasks. Unlike its "Flash" counterparts that use Mixture-of-Experts (MoE) architectures, the 27B variant is a dense model, providing consistent performance and high reliability for complex multi-step reasoning.

The model utilizes a hybrid-attention architecture, incorporating linear attention on 48 of its 64 layers to optimize performance during long-context processing. It supports a native context window of 262,144 tokens, which can be extended up to one million tokens in hosted environments. As a native vision-language model, it natively understands images and video content, ranging from technical STEM diagrams to hour-scale video files, without requiring separate vision encoders.

Capabilities and Thinking Control

A key feature of Qwen3.8 27B is "Flexible Thinking Control," which allows users to toggle between reasoning-heavy "Thinking" modes and a standard "Non-reasoning" mode. In the non-reasoning mode, the model functions as a high-speed causal LLM, bypassing the internal chain-of-thought overhead to significantly reduce latency and token usage. This mode is particularly suited for high-throughput production workflows, standard dialogue, and environments where reasoning is handled by an external agentic framework.

In addition to its text and vision capabilities, the model is highly optimized for tool use and autonomous planning. It exhibits strong performance in handling environment feedback, making it effective for end-to-end task completion in developer environments and professional office suites. The model's weights are released under a permissive license, facilitating local deployment on consumer-grade hardware with high-capacity unified memory or multi-GPU setups.