阿里巴巴Qwen团队发布了Qwen3.8-27B开源模型，这是Qwen系列中参数为27B的原生多模态密集模型[1][3]。该模型性能超越Qwen3.7-Plus[3]，采用Apache 2.0许可证[3]，可在Hugging Face和ModelScope平台下载[3]。
Qwen3.8-27B具备较强的技术规格[1]。模型包含64层隐藏层，隐藏维度为5120[1]，原生上下文长度为262,144个token，可通过YaRN技术扩展至1,000,000个token[1][3]。模型内置灵活的思考模式控制，支持xhigh（默认）、medium、low三个推理effort等级[1]，思考模式下推荐参数为temperature=1.0、top_p=0.95、top_k=20[1][2]；指令模式推荐参数为temperature=0.7、top_p=0.80、top_k=20[2]。
该模型支持原生视觉-语言理解和长文本处理能力[1][2]，在编程、专业工作、研究和智能体任务上有实质性提升[1]。Qwen3.8-27B兼容多个主流推理框架，包括Hugging Face Transformers、vLLM、SGLang和TokenSpeed[1]，同时提供了FP8量化版本便于本地部署[1]。Unsloth还提供了基于该模型的GGUF量化版本及完整使用指南[2]，支持llama.cpp、Ollama、LM Studio等推理工具[2]。
此外，Qwen团队同时发布了更大规模的Qwen3.8-2.4T-A95B（Max-level）模型的开源权重[3]。
Alibaba's Qwen team has released Qwen3.8-27B, an open-source dense model with 27 billion parameters that represents the strongest version in the Qwen open-source lineup [1][3]. The model features native multimodal capabilities including image and video understanding, supports a native context length of 262,144 tokens that can be extended to 1 million tokens [1][3], and is distributed under the Apache 2.0 license [3].
The 64-layer model with 5,120 hidden dimensions delivers substantial improvements in programming, professional work, research, and agentic tasks [1]. It incorporates flexible thinking mode controls with configurable inference effort levels—xhigh (default), medium, and low—allowing users to balance reasoning depth against computational cost [1]. The default thinking mode parameters are set to temperature 1.0, top_p 0.95, and top_k 20, while instruction mode uses temperature 0.7, top_p 0.80, and top_k 20 [2]. The model performs better than its predecessor Qwen3.7-Plus [3].
To facilitate local deployment, the Qwen team provides fine-grained FP8 quantization with a block size of 128 [1], and Unsloth has released GGUF quantized versions with integration support for frameworks including llama.cpp, Ollama, and LM Studio [2]. The base model is compatible with Hugging Face Transformers, vLLM, SGLang, and TokenSpeed [1]. Additionally, Alibaba has released the open weights for the larger Qwen3.8-2.4T-A95B model [3], with both models available for download on Hugging Face and ModelScope [3].