2026年8月26日，阿里巴巴发布了开源多模态推理模型Qwen3.8-Flash-Next 1。在Artificial Analysis Intelligence Index评估中，该模型得分56分，显著高于同类开源模型28分的中位数 1。该模型的输入与输出价格均为每百万tokens 0.00美元，其开源权重采用Qwen Community License 1.0许可证 1。
在技术架构方面，Qwen3.8-Flash-Next采用混合专家（MoE）架构，总参数为1800亿，活跃参数为60亿 1。该模型支持文本、图像和视频输入以及文本输出，上下文窗口为260k tokens 1。在性能表现上，其输出速度达73.5 tokens/s（同类中位数为65.1 tokens/s），首token延迟（TTFT）为2.72秒（同类中位数为2.27秒） 1。此外，在评估阶段，该模型共生成了2亿个输出tokens，高于1.1亿个的中位数水平 1。
Alibaba released Qwen3.8-Flash-Next, an open-source multimodal reasoning model, on August 26, 2026 1. Built on a Mixture of Experts (MoE) architecture, the model comprises 180 billion total parameters with 6 billion active parameters 1. It is capable of processing text, image, and video inputs to generate text outputs, and it supports a context window of 260,000 tokens 1.
Evaluations reveal that Qwen3.8-Flash-Next scored 56 on the Artificial Analysis Intelligence Index, substantially exceeding the median score of 28 for comparable open-source models 1. The system achieved an output speed of 73.5 tokens per second against a median of 65.1 tokens per second, and it generated 200 million output tokens during testing, well above the median of 110 million 1. Its time to first token (TTFT) was measured at 2.72 seconds, slightly trailing the median of 2.27 seconds 1. The model's weights are available under the Qwen Community License 1.0, with both input and output pricing set at $0.00 per 1 million tokens 1.
评论
还没有评论，欢迎留下第一条。