From the source
Thinking Machines Lab has released Inkling 2 Mini, a looped-transformer multimodal Mixture-of-Experts model built for local AI setups, with about 26B physical and 3B active parameters per token.
- The model has 24 decoder layers, but its 12 middle layers run three times in a loop, so it behaves like 48 logical layers while storing only 24 physical ones.
- It supports a 1M-token context and accepts images and text, and ships in a mixed-precision format with NVFP4 and MXFP8 expert weights plus a DSpark drafter for extra speed at a small memory cost.
Hugging Face says the model is state-of-the-art in its class on the Artificial Analysis Intelligence Index v5 and leads in 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL.
- In vibe tests it was at least comparable with Qwen 3.8 27B, though the post says more validation is needed.
- Transformers, llama.cpp and MLX support are provided, and SGLang offers day-zero inference support for production deployments on datacenter hardware.
- A separate research artifact, TML-d24-varloop-research, shares the architecture but accepts a variable number of loop passes; setting 1 or 2 passes instead of 3 speeds inference considerably, with possibly slightly lower quality.






