Thinking Machines releases Inkling 2 Mini, an open-weights looped transformer that runs on a laptop

The 26.4B-parameter mixture-of-experts model runs its middle layers three times per token, uses about 20GB, and the lab says it leads open models under 100B parameters on DeepSWE v1.1 and Vals Index v2.2.
Why it matters
Most small open models trade quality for size; Thinking Machines instead reuses its middle layers, so the model reasons more per token without writing longer chains of thought. The lab says this matters for monitoring, and reports that a variable-loop research variant kept a legible chain of thought and did not score higher on agentic-sabotage tests as loops were added. ARC Prize lists 49.3% on ARC-AGI-2. Watch independent coding evaluations and whether other labs adopt looping for on-device models.
Key facts
- 26.4B total parameters, 3.4B active
- Middle layers run three times per token
- About 20GB footprint, mostly NVFP4
- DeepSWE v1.1: 48.2%; Vals Index v2.2: 46.5%
- ARC-AGI-2: 49.3% (ARC Prize)
- Variable-loop research model released too