ModelsMulti-source

Thinking Machines 开源 Inkling 2 Mini:循环推理省 token,笔记本可跑

New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop

Why we picked it

Thinking Machines 放出的 Inkling 2 Mini 用循环重复 12 层的办法做内部推理,不靠长思维链,成本只有 Qwen3.8 27B 的约六分之一还打平成绩。想在笔记本上本地跑推理模型的开发者可以重点关注这个架构。

Thinking Machines 发布开源模型 Inkling 2 Mini,采用循环 transformer 架构,24 层中让 12 层对每个 token 重复执行三次,从而在内部延长推理过程而不产生额外输出文本。

Body

在 Artificial Analysis 的 Intelligence Index 基准上,它与 Qwen3.8 27B 打平,成本约低六倍。

由于总参数规模较小,该模型可以在普通笔记本上本地运行。循环机制的含义是把算力花在重复利用层参数上,而不是堆叠更深的网络或生成更长的思维链,为低算力场景提供了一条不同的推理路径。

对开发者来说,这意味着在不依赖云端 GPU 的情况下获得接近更大开源模型的推理表现,同时显著降低 token 开销。

Original · Decoder

New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop

Thinking Machines has released Inkling 2 Mini, an open looped transformer that runs 12 of its 24 layers three times per token. This lets the model reason internally for longer without generating extra text. On Artificial Analysis's Intelligence Index, it ties with Qwen3.8 27B at roughly six times lower cost. The article New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop appeared first on The Decoder .

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. 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 Inkling 2 Mini, Thinking Machines, Qwen3.8, open-source models.

Source attribution: curated from Decoder (@Matthias Bastian). Cross-source corroboration: 11 independent sources have covered this development (including Artificial Analysis, Julien Chaumond, kimmonismus), confirming industry-wide signal strength.

  • Artificial Analysis11-20 20:35Source
  • Julien Chaumond11-20 20:18Source
  • kimmonismus11-20 20:22Source
  • lmarena.ai11-20 22:58Source
  • Clement Delangue11-20 20:17Source
  • Merve Noyan11-20 20:24Source
  • Thinking Machines Lab11-20 20:10Source
  • vLLM11-20 20:14Source
  • ollama11-20 21:21Source
  • Mira Murati (TML)11-20 20:14Source