阿里巴巴Qwen团队发布了Qwen3.8-2.4T-A95B大语言模型，这是Qwen开源系列中首个Max级别的模型[1]。该模型具有2.4T总参数和95B激活参数，采用92层架构、512个专家模块的设计，每次激活10个路由专家与1个共享专家[1]。在代码、专业工作、研究和长链任务等领域相比之前版本有显著提升[1]。
模型在上下文处理能力上表现突出，原生支持262,144个tokens的上下文长度，可灵活扩展至1,010,000个tokens[1]。为满足不同应用场景，该模型支持reasoning_effort参数调节推理深度，提供xhigh（默认）、medium和low三个级别的选项[1]。推理内容最大输出长度建议为262,144个tokens，最终响应长度建议不超过131,072个tokens[1]。采样参数方面，官方推荐设置temperature=1.0、top_p=0.95、top_k=20、min_p=0.0[1]。该模型兼容vLLM、SGLang、TokenSpeed等多个推理框架[1]。
Alibaba's Qwen team has unveiled Qwen3.8-2.4T-A95B, marking the most capable version in the Qwen open-source model series and the first Max-tier model made publicly available [1]. The model features 2.4 trillion total parameters with 95 billion active parameters and demonstrates significant improvements in code generation, professional applications, research tasks, and extended-context reasoning [1].
The model operates with 92 layers and 512 expert modules, activating 10 routed experts plus one shared expert per inference step [1]. It supports a native context length of 262,144 tokens, expandable to 1,010,000 tokens, enabling substantially longer document processing and reasoning chains [1]. Users can control inference depth through the reasoning_effort parameter, which offers three settings—xhigh (default), medium, and low—providing flexibility for different computational and accuracy trade-offs [1].
Recommended inference parameters include a temperature setting of 1.0, top_p of 0.95, top_k of 20, and min_p of 0.0 [1]. The model supports maximum reasoning content output of 262,144 tokens and final response output of 131,072 tokens [1]. Integration is available with multiple inference frameworks including vLLM, SGLang, and TokenSpeed [1].