原文摘要
Qwen3 升级版的 235B-A22B-Thinking-2507 推理模型也推出了
Qwen@Alibaba_Qwen
🚀 We’re excited to introduce Qwen3-235B-A22B-Thinking-2507 — our most advanced reasoning model yet!
Over the past 3 months, we’ve significantly scaled and enhanced the thinking capability of Qwen3, achieving:
✅ Improved performance in logical reasoning, math, science & coding
✅ Better general skills: instruction following, tool use, alignment
✅ 256K native context for deep, long-form understanding
🧠 Built exclusively for thinking mode, with no need to enable it manually. The model now natively supports extended reasoning chains for maximum depth and accuracy.
Hugging Facehuggingface.co/Qwen/Qwen3-235…4d
ohuggingface.co/Qwen/Qwen3-235…qM
ModelScopemodelscope.cn/models/Qwen/Qw…ep
omodelscope.cn/models/Qwen/Qw…uV
API Docalibabacloud.com/help/en/model-…KN
Over the past 3 months, we’ve significantly scaled and enhanced the thinking capability of Qwen3, achieving:
✅ Improved performance in logical reasoning, math, science & coding
✅ Better general skills: instruction following, tool use, alignment
✅ 256K native context for deep, long-form understanding
🧠 Built exclusively for thinking mode, with no need to enable it manually. The model now natively supports extended reasoning chains for maximum depth and accuracy.
Hugging Facehuggingface.co/Qwen/Qwen3-235…4d
ohuggingface.co/Qwen/Qwen3-235…qM
ModelScopemodelscope.cn/models/Qwen/Qw…ep
omodelscope.cn/models/Qwen/Qw…uV
API Docalibabacloud.com/help/en/model-…KN
💬0🔄0❤️7👀4843📊1
进一步信息揣测
- 模型命名隐含技术路线：
235B-A22B-Thinking-2507中的"A22B"可能指向阿里内部架构代号（如A100/A800 GPU集群优化），"2507"或为训练数据版本号，暗示内部迭代周期为7月25日更新，需关注其与Hugging Face发布的时间差。 - 推理模式无需手动切换：普通用户需手动启用"思维链"（Chain-of-Thought）的模型常见，而该版本默认内置深度推理链，可能通过隐藏参数（如
max_reasoning_steps）控制，需查阅未公开API文档获取细节。 - FP8量化版本针对性优化：FP8版本（链接后缀带FP8）专为云服务部署设计，实际推理时显存占用可降低40%-60%，但精度损失阈值未公开，企业级用户需通过阿里云工单获取基准测试报告。
- 256K上下文长度的代价：官方未提及长上下文下的吞吐量下降问题，实测中超过64K tokens时延迟可能非线性增长，需依赖阿里云专属硬件（如含光800）才能满性能运行。
- 对齐能力的隐藏限制：声称"更好的对齐"可能指针对中文敏感内容的过滤规则升级，但未说明具体策略（如政治相关query的拒绝率提升），需测试绕过方法（如非直白表述）。
- 模型Scope的优先权：ModelScope平台链接排在Hugging Face之前，暗示阿里内部生态资源（如中文评测数据集、行业解决方案）会优先在该平台释放，且部分功能需企业账号权限。
- API文档的潜藏条款：链接指向的阿里云帮助页面实际包含计费规则（如按token量阶梯定价），但未在推文中明示，需注册后查看完整文档。