Inkling 2 Mini
Intro
Inkling 2 Mini is an open-weights model that Thinking Machines Lab released on 20 November 2026. It is a smaller member of the Inkling 2 family, distinct from the already listed Inkling 2 Small (594B total parameters), and the lab says it is small enough to run on a laptop. Official announcement: Introducing Inkling 2 Mini. The official account @thinkymachines posted at 2026-11-20T20:10:34Z (04:10 Beijing time on 21 November): Some of you thought that Inkling 2 Small was too big….
The official page describes a Mixture-of-Experts looped transformer whose middle layers run three times per token, so the model can do more internal reasoning while writing fewer tokens. It has 26.4B total parameters and 3.4B active parameters per token, or about 5.4B weight activations per token once looping is counted. The lab says it leads open-weights models with under 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%); the official post also cites 20.1 on the Artificial Analysis Index. A variable-loop research variant is released alongside it for monitorability experiments. There is also a Hugging Face blog post: https://huggingface.co/blog/inkling2mini .
Advertisement
Explain
What problem it targets
Inkling 2 Small has 594B total parameters, which is still large for people who want a local setup or who want to study looping itself. Inkling 2 Mini puts a looped transformer in a 26B-class model so more people can run it on a laptop, and so the safety and monitorability community can study how loops affect chain-of-thought supervision.
Technical points from the official page
- Looped MoE: the middle layers run three times per token, trading repeated compute for a shorter but still legible chain of thought;
- Scale: 26.4B total parameters, 3.4B active, and about 5.4B weight activations per token;
- Training and serving: the lab says most of the weights were trained and are deployed natively in NVFP4;
- Variable-loop research variant: used to compare loop counts. The official write-up says the no-CoT time horizon improved from 16 seconds at one pass to 37 seconds with three; serial-reasoning performance kept improving out to ten loops, twice as many as it was trained with; the chain of thought grew shorter but stayed legible; and tests of agentic sabotage did not trend higher as loops were added.
Limits
The official comparisons are against open-weights models with under 100B total parameters, and evaluation setups can vary by model — the lab says so. This is not simply a shrunk copy of Inkling 2 Small: the two models differ by an order of magnitude in size and in intended use. This entry does not treat unofficial memory-footprint figures or third-party recaps as official numbers.
Usage
Who it is for
- People who want an open-weights model small enough for a laptop or a lighter agentic-coding setup;
- Safety and alignment researchers studying looped architectures and chain-of-thought monitorability;
- Teams already following Inkling 2 Small who need a much smaller checkpoint.
How to use it
- Read the official announcement: https://thinkingmachines.ai/news/inkling-2-mini/ ;
- Official post: https://x.com/thinkymachines/status/2123871643797995872 ;
- There is also a Hugging Face write-up: https://huggingface.co/blog/inkling2mini ;
- When you quote comparisons, keep the official frame: open-weights models under 100B total parameters, plus DeepSWE v1.1, Vals Index v2.2, and the official post’s AA Index score of 20.1.
Serving entry points, the weight license, and download details should be taken from the official site and Hugging Face at the time you use them; this entry does not infer those details.