Thinking Machines Releases Inkling 2 Mini, a 26B Looped Transformer That Runs on Laptops
Thinking Machines Lab (Mira Murati's team) released the open-weights model Inkling 2 Mini on November 21: a MoE "looped transformer" with 26.4B total parameters but only 3.4B activated per token, small enough to run on a laptop, and it has topped small-model coding leaderboards. Its core innovation is letting the model "think longer" without writing out its reasoning.
Confirmed
- Architecture: The model has 24 layers, with the middle 12 layers executed three times per token, while the first and last 6 layers run only once; the looping mechanism lets the model "reason more while writing less."
- Monitorability findings: The team reported that as loop count increases, no-CoT (not writing out chain-of-thought) ability rises; at this scale, chains of thought become shorter yet remain readable, with no noticeable drop in monitorability. The team also released a variable-loop research model for the community to verify independently.
- Availability: Runs on a 32GB Mac via Ollama, supports NVIDIA DGX Spark, and is available on Tinker for inference and fine-tuning.
Unconfirmed
- The team acknowledged in the release that the looped architecture raises monitorability concerns; while no clear degradation was observed in these experiments, whether this holds at larger scales or in other settings remains an open question.
Why it matters
- Looped transformers offer an "implicit deliberation" route distinct from long chain-of-thought: compute scales with loop count while parameter count stays fixed, giving small models a low-cost way to boost reasoning.
- With open weights, local execution, and fine-tuning support, the community can easily reproduce it and build on it for research on monitorability, variable loops, and related directions.
2026-11-21 ~ 2026-11-21 · 10 related posts
Primary sources
- Thinking Machines releases Inkling 2 Mini, a 26B looped transformer that runs on laptops — thinkymachines ·
- Inkling 2 Mini architecture: 24 layers, with the middle 12 looped three times — thinkymachines ·
- Thinking Machines: more loop passes boost no-CoT capability without hurting monitorability — thinkymachines ·
- Thinking Machines releases Inkling 2 Mini, a looped transformer topping sub-100B open-weights coding benchmarks — SonglinYang4 · 2026-11-21
- [source] Inkling 2 Mini architecture: 24 layers, with the middle 12 looped three times — thinkymachines · 2026-11-21
- [source] Thinking Machines: more loop passes boost no-CoT capability without hurting monitorability — thinkymachines · 2026-11-21
- Inkling 2 Mini runs on 32GB Macs via Ollama and on NVIDIA DGX Spark — thinkymachines · 2026-11-21
- Thinking Machines releases Inkling 2 Mini, a 26B looped transformer that runs on laptops — miramurati · 2026-11-21
- Thinky Machines releases open-weights Inkling 2 Mini, a 26B looped transformer for safety research — SonglinYang4 · 2026-11-21
- Thinking Machines ships Inkling 2 Mini: 26B MoE that runs on 16GB RAM, tops its size class — SonglinYang4 · 2026-11-21
- Thinking Machines Lab ships Inkling 2 Mini: looped MoE with 3B active params for local setups — ariG23498 · 2026-11-21
2 near-duplicate retellings: thinkymachines · victormustar