New models Thinking Machines Lab

Thinking Machines Lab releases Inkling 2 Mini, a looped mixture-of-experts model for local use

Thinking Machines Lab has released Inkling 2 Mini, a smaller sibling of Inkling 2 Small aimed at running on local machines. It is a multimodal mixture-of-experts model that takes text and images, handles a context of up to 1M tokens and has about 26B total parameters with roughly 3B active per token. Its trick is a loop: the 12 middle decoder layers run three times, so 24 physical layers behave like 48. Weights ship in mixed NVFP4 and MXFP8 precision with a DSpark drafter for speculative decoding. Hugging Face reports day-one support in transformers, llama.cpp, MLX and SGLang, and says early tests put it on par with Qwen 3.8 27B.

Source

  1. Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup (opens in a new tab) Hugging Face Official ·

How we write topics: our editors read the official announcement and other reliable reporting, then write a short neutral summary in their own words. We link to the original sources and do not republish them. Spotted a mistake? Tell us. ← All AI news

Emoji artwork: Twemoji (CC BY 4.0)