United StatesThinking Machines’ Inkling 2 Mini runs on a laptop in 20GB.
Its makers warn looping lets models think more, output less and may become harder to monitor.
The middle 12 of its 24 layers run 3 times per token, so it finds answers with fewer written tokens.
It trails Qwen3.8-27B on multi-turn chat and multi-constraint instructions.
CEO Mira Murati also released a variable-loop research model so outsiders can test monitorability.
How each outlet framed it
- Zhidx leans favorable
- credits the model's recurrent architecture, which runs middle layers three times per token, for fewer output tokens and a laptop-scale footprint, and notes Qwen4-27B's imminent release as a challenge
Sources: Zhidx