Thinking Machines开源多模态模型Inkling 2 Mini,可在笔记本本地运行
Thinking Machines Lab于2026年11月21日发布开源权重多模态模型Inkling 2 Mini。该模型总参数26.4B,内存占用约20GB,可在32GB MacBook Pro或单张桌面GPU上离线运行;权重以Apache 2.0许可在Hugging Face发布,并附带模型可接受使用政策。
Inkling 2 Mini采用循环Transformer架构:24层解码器稀疏MoE,每个token路由至128个专家中的8个,中间12层重复执行三次,等效48层应用、5.4B有效激活参数。官方称其在软件工程与工具调用基准上领先同规模开源模型,综合智能指数20.1与阿里Qwen3.8-27B持平;Artificial Analysis称其每任务速度快约4倍、成本低约6倍,计算量少约2.8倍。
该模型原生支持文本与图像输入,上下文窗口100万token。基准上,DeepSWE v1.1为48.2%,SWE-bench Verified为79.2%。不过其AA-Omniscience v1.1事实召回得分-8.6,低于Qwen3.8-27B;多轮聊天与多约束指令测试也落后,非英语中文偶有截断。Thinking Machines还发布可变循环检查点TML-d24-varloop-research,与NDIF、Goodfire研究可解释性。
- 2026-11-22 02:29 | venturebeat.com:Thinking Machines wows with Inkling 2 Mini looped transformer model that runs locally on laptops阅读原文
Thinking Machines Lab, the San Francisco-based AI startup started and led by former OpenAI chief technology officer Mira Murati, is challenging the dominant paradigm of cloud-based AI frontier models with the release yesterday of Inkling 2 Mini, an open-weights multimodal model engineered to run off...