Event ·Multi-source

Thinking Machines Lab开源Inkling 2 Mini:26.4B参数可跑在16GB内存笔记本上

前OpenAI CTO Mira Murati的团队Thinking Machines Lab开源了Inkling 2 Mini,采用循环Transformer架构,总参数26.4B、激活参数仅3.4B,可在16GB内存的笔记本上离线运行。其核心机制是改变Transformer内部的计算方式:通常每个token要过一遍所有层,而Inkling 2 Mini让中间12层循环跑3遍,用24层权重干出48层的效果。对照实验显示,不循环版本需要多花140%推理算力才能追平它的表现。 基准成绩方面,Inkling 2 Mini在SWE-bench Verified拿到79.2%,DeepSWE编程任务48.2%,BrowseComp 79%,领先同规模开源模型。研究团队还一并开源了可变循环次数的研究版模型TML-d24-varloop,用于监控循环次数变化时模型的思考过程。当主流路线都在堆参数时,这条“堆层数使用次数”的scaling路径提供了一个不同的方向。

Current read

Mira Murati团队开源的Inkling 2 Mini用循环层数的思路把24层权重跑出48层效果,激活参数只有3.4B,普通笔记本就能离线跑。SWE-bench Verified 79.2%的成绩领先同规模开源模型,想研究推理效率或本地部署的开发者可以下载试试。

11 sources79° AI heatLast updated 2026/11/21 10:36:52

Evidence

11 sources
Related source 1kimmonismus
View source
Related source 3marktechpost
View source
Related source 4Julien Chaumond
View source
Related source 5Artificial Analysis
View source
Related source 6Merve Noyan
View source
Related source 7Clement Delangue
View source
Related source 8lmarena.ai
View source
Related source 9TestingCatalog
View source
Related source 10Thinking Machines Lab
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Thinking Machines Lab开源Inkling 2 Mini:26.4B参数可跑在16GB内存笔记本上 · 事件专题 | AITOP