A 26B model that loops its middle 12 layers three times: Thinking Machines' Inkling 2 Mini runs on a laptop
First seen on X 7 hours ago@thinkymachines ♥ 4,2257 views
Mira Murati's Thinking Machines Lab released Inkling 2 Mini, an open-weights model small enough to run on a laptop.
"Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you," the company posted on Nov. 20, drawing 4,300 likes. Inkling 2 Mini is a looped transformer: of its 24 layers, the middle 12 run three times per token while the first and last six run once, so the model reasons more while writing less. It has 26.4B total parameters with 5.4B active per token once looping is counted, was trained and ships natively in NVFP4 at around 20GB, and is licensed Apache 2.0.
You can try it now. Weights are on Hugging Face, Ollama supports it in the 0.40.17 pre-release on Macs with 32GB or more, and Unsloth published GGUF files it says run in 16GB of RAM. It also runs on Nvidia's DGX Spark and is available on Thinking Machines' Tinker for inference and fine-tuning.
The company says the model leads every open model under 100B parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). Artificial Analysis scores it 20 on its Intelligence Index, tied with Qwen3.8-27B, at $0.59 and 9.0 minutes per task versus $3.53 and 38.2 minutes for Qwen. GPT-6 Luna, though, scores 28 for $0.15 per task, and Thinking Machines says Mini trails Qwen on multi-turn chat and heavily constrained instructions.
Looping also raises a safety question, since a model can think longer without writing its reasoning down. Thinking Machines says the chain of thought stayed legible at this scale but "this may not stay the case," and released a variable-loop research model; Goodfire got early access to study its internals.
Comments 0
Select a sentence in the story (long-press on phones) to comment on it.
“Some of you thought that Inkling 2 Small was too big at 594B parameters.”
