Qwen 4 Max API 上线,CoT 长度明显缩短
Qwen 4 Max API 刚上线,实测亮点是 CoT 更短、单位 token 智能密度更高。如果你在等能本地部署的小模型,作者点名的新 27B 变体值得盯,做领域微调和私有化部署的人尤其应该关注后续发布。
Qwen 4 Max API 正式上线。博主 Xianbao_QIAN 试用后称其在智能密度上有大幅提升,尤其是 CoT(思维链)长度显著减少,推断每步所需的推理 token 更少。
帖子中没有给出具体基准分数,也不确定 CoT 缩减来自哪种训练方法。作者预期 smaller 变体采用同样的数据和方法训练,并特别关注即将推出的新 27B 版本:该尺寸适合做领域微调和本地部署。
对开发者来说,这条消息的信号是 Qwen 系列继续走"大模型能力下沉到小尺寸"的路线,27B 若确认同源训练,会是本地部署场景里值得对比测试的选项。
Qwen 4 Max API is out.
It seems to have a huge intelligence density win. Not sure where this CoT reduction capability is coming from but if the smaller variants to be released soon is trained on the same data / method, I'll be really looking for the new 27B variant that can be turbo charged with domain adaption and local deployment!
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Qwen, Qwen4 Max, open-source models, reasoning models.
Source attribution: curated from Xianbao Qian (@Xianbao_QIAN). Cross-source corroboration: 11 independent sources have covered this development (including 阿里通义 Qwen, 阿里云 Alibaba Cloud, Decoder), confirming industry-wide signal strength.