Qwen Unveils Qwen3.8-27B: A Dense Vision-Language Model with FP8 Quantisation
Built on the Qwen3.5 architecture, the new 27-billion parameter model introduces native support for images and videos, configurable thinking modes, and compatibility with major inference frameworks.
The Qwen team has released Qwen3.8-27B, a dense vision-language model that builds upon the Qwen3.5 architecture. Positioned as the most capable generation in the Qwen open-model family to date, the model delivers substantial gains in coding, professional work, research, and long-horizon agentic tasks. The release includes FP8-quantized model weights and configuration files in the Hugging Face Transformers format, designed for streamlined integration into existing infrastructure.
The model features native support for images and videos, alongside flexible thinking control mechanisms. Qwen3.8-27B operates in 'thinking mode' by default, generating internal reasoning content signified by specific tokens before producing a final response. Users can adjust the depth of this reasoning and associated costs using the 'reasoning_effort' parameter. Additionally, the 'preserve_thinking' feature is enabled by default to maintain context continuity across multi-turn conversations, which is particularly beneficial for agent scenarios requiring decision consistency.
In terms of capacity, the model supports up to 262,144 tokens natively. For long-horizon tasks exceeding this limit, the Qwen team recommends using RoPE scaling techniques such as YaRN, which are supported by frameworks including vLLM, SGLang, and TokenSpeed. To optimise video understanding, users are advised to set the 'longest_edge' parameter in the video_preprocessor_config to 469,762,048, enabling higher frame-rate sampling for hour-scale videos.
The quantization method utilises fine-grained FP8 with a block size of 128. According to the release details, performance metrics are nearly identical to the original model, ensuring that efficiency gains do not come at the cost of accuracy. The model is compatible with popular inference frameworks, including vLLM, SGLang, TokenSpeed, and Hugging Face Transformers, with the team recommending the latest framework versions for optimal performance and compatibility.
For organisations seeking managed, scalable inference without the burden of infrastructure maintenance, the official Qwen API service is provided via Qwen Cloud. This hosted version will offer production features, including a 1M context length by default and built-in tools. The service is described as 'coming soon', with further details available in the Qwen3.8-27B Overview documentation.
