Thinking Machines releases Inkling 2 Mini, a 26B looped transformer that runs on laptops
Thinking Machines Lab (founded by Mira Murati) released the open-weights small model Inkling 2 Mini on November 21: a 26.4B-total-parameter MoE "looping transformer" with roughly 3.4B parameters activated per token, directly responding to earlier user feedback that the 594B-parameter Inkling 2 Small was too large. Small enough to run locally on a laptop, it became the talk of the open-source community that day.
Confirmed
- Architecture: 24 layers total, with the middle 12 layers executed three times in a loop per token while the 6 layers at each end run only once; the team says looping lets the model "reason more, write less." Multimodal (image-text-to-text) with 1M context support
- Availability: Hugging Face (thinkingmachines/Inkling-2-Mini, which topped the trending list), Ollama (v0.40.17 pre-release, ollama run inkling-2-mini), llama.cpp/Llama platform on day one, plus Tinker for inference and fine-tuning; the team says it runs on a 32GB-memory Mac, with NVIDIA DGX Spark also supported—multiple posts note around 20GB of memory is enough
- License: Apache 2.0, no subscription or account needed, works offline; Unsloth has already released GGUF quantized versions
- Monitorability research: the team reports that as loop counts increase, no-CoT capability (without chain-of-thought) rises; at this scale chains of thought get shorter but remain readable, with no clear drop in monitorability—they also released a variable-loop research model for the community to verify
- Performance: per poster @victormustar, the model leads same-size models on DeepSWE and SWE-bench Pro (the post did not include full benchmark comparison details)
Unconfirmed
- Posts disagree on activated parameters: the official post says 3.4B while @WorldofAI's hands-on post says 5.4B—the model card should be authoritative
- @alexverem called it "the first free competitor to OpenAI's cheap models," which is the author's own framing and not an official claim
Why it matters
Looping transformers move "more reasoning" inside the model itself instead of generating longer chains of thought, directly touching on monitorability—a key safety concern; meanwhile, the 26B scale, Apache 2.0 license, and laptop-runnable deployment bar make it a landmark release for local/offline open-source models, with Ollama, llama.cpp, Unsloth, and the broader ecosystem adding support on day one.
2026-11-20 ~ 2026-11-22 · 111 related posts
Primary sources
- Thinking Machines releases Inkling 2 Mini, a 26B looped transformer that runs on laptops — thinkymachines ·
- Thinking Machines releases Inkling 2 Mini, a 26B looped transformer that runs on laptops — miramurati ·
- Thinking Machines' Inkling 2 Mini lands on Ollama, runs locally on a 32GB Mac — ollama ·
- Unsloth ships GGUF quantization of Thinking Machines' Inkling-2-Mini multimodal MoE — unsloth · 2026-11-20
- Thinking Machines releases Inkling 2 Mini, a looped transformer topping sub-100B open-weights coding benchmarks — SonglinYang4 · 2026-11-21
- Inkling 2 Mini architecture: 24 layers, with the middle 12 looped three times — thinkymachines · 2026-11-21
- 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 open-sources Inkling-2-Mini: looped MoE with 26B params, only ~5.4B effective active — Skyline34rGt · 2026-11-21
- [source] 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
- Inkling 2 Mini runs locally in 16GB RAM: 5.4B-active MoE tops its size class — simonguozirui · 2026-11-21
- unsloth releases GGUF quants of Inkling-2-Mini for local inference — coder543 · 2026-11-21
- Thinking Machines Lab ships Inkling 2 Mini: looped MoE with 3B active params for local setups — ariG23498 · 2026-11-21
- Thinking Machines Ships Inkling 2 Mini: 26B Looped Transformer That Runs on Laptops — altryne · 2026-11-21
- Thinking Machines open-sources Inkling 2 Mini: 26B looped transformer with 5.4B active params — SonglinYang4 · 2026-11-21
- Inkling 2 Mini: a 26B looped MoE that runs agentic workloads locally at 168 tok/s — ying11231 · 2026-11-21
- Thinking Machines ships Inkling 2 Mini: 26B looped transformer that runs on laptops — LiTianleli · 2026-11-21
- Thinking Machines open-sources 26B variable-loop model; Inkling 2 Mini beats 594B predecessor at 2.7% size — ChengleiSi · 2026-11-21
- Thinky Machines ships Inkling 2 Mini: 26B looped MoE that runs on 32GB machines at 75 tok/s — simonguozirui · 2026-11-21
- Thinking Machines ships Inkling 2 Mini: 26B looped transformer that runs on your laptop — simonguozirui · 2026-11-21
- Inkling 2 Mini runs a local Mac assistant: 26.4B params, only 3.4B active — simonguozirui · 2026-11-21
- Thinking Machines' Inkling-2-Mini trends on Hugging Face — thinkingmachines · 2026-11-21
- [source] Thinking Machines' Inkling 2 Mini lands on Ollama, runs locally on a 32GB Mac — ollama · 2026-11-21
- Inkling 2 Mini on Ollama: MLX build, one-command hooks into Claude Code and OpenCode — ollama · 2026-11-21
- Thinking Machines quietly releases Inkling 2 Mini, details on official site — shadowt1tan · 2026-11-21
- 26B Looped Transformer Inkling 2 Mini Runs on a Laptop, Scores 68.4% on BVB Video Benchmark — SonglinYang4 · 2026-11-21
- Mira Murati's Thinking Machines releases Inkling 2 Mini, a looped transformer that runs on a laptop — simonguozirui · 2026-11-21
- Thinking Machines' 26B looped-MoE Inklings 2 Mini beats GLM-5.2 and Opus 4.5 on benchmarks — ziqiao_ma · 2026-11-21
- 'Laptop-sized but agentic-capable': Inkling 2 Mini's 26B looped transformer wows developers — LiTianleli · 2026-11-21
- Thinking Machines ships Inkling 2 Mini: 26B looped MoE with NVFP4-native weights, ~20GB for laptops — testingcatalog · 2026-11-21
- Inkling 2 Mini benchmarks: 26B looped transformer beats same-size rivals on SWE tasks — clarejtbirch · 2026-11-21
20 near-duplicate retellings: thinkymachines · victormustar · mervenoyann · himanshustwts · ziqiao_ma · pcuenq · SonglinYang4 · simonguozirui · DimitrisPapail · mark_k · SonglinYang4 · cogconfluence · clarejtbirch · keirp1 · LiTianleli · ChowdhuryNeil · LiTianleli · ziqiao_ma · simonguozirui · DavideCrapis