阿里巴巴Qwen团队推出了Qwen3.8-2.4T大语言模型的FP8量化版本[1]，这是Qwen开源模型家族中首次发布Qwen-Max级别的开源模型[1]。该模型拥有2.4T总参数和95B激活参数[1]，采用细粒度FP8量化方法，块大小为128[1]，性能指标与原始模型基本相同[1]。
模型在部署和推理上具有多项特性[1]。原生上下文长度支持262,144令牌，可扩展至1,010,000令牌[1]。推理深度方面，模型支持通过reasoning_effort参数进行调整，选项包括xhigh、medium和low[1]。默认启用preserve_thinking功能，以保留历史消息的推理上下文[1]。该模型支持vLLM、SGLang、TokenSpeed等多种推理框架部署[1]。建议的采样参数配置为temperature=1.0、top_p=0.95、top_k=20[1]。
Alibaba has unveiled the FP8-quantized version of Qwen3.8-2.4T, marking the first time the Qwen open-source model family has introduced a model at the Qwen-Max performance level [1]. The model features 2.4 trillion total parameters with 95 billion active parameters, utilizing fine-grained FP8 quantization with a block size of 128 [1]. Performance metrics remain comparable to the original model despite the quantization approach [1].
The model supports native context lengths of 262,144 tokens, which can be extended to 1,010,000 tokens [1]. It is compatible with multiple inference frameworks including vLLM, SGLang, and TokenSpeed [1]. The model includes a reasoning_effort parameter that allows users to adjust inference depth across three settings: xhigh, medium, and low [1]. By default, the preserve_thinking feature is enabled, maintaining reasoning context from historical messages [1]. For optimal performance, the recommended sampling parameters are temperature set to 1.0, top_p at 0.95, and top_k at 20 [1].