Hugging Face released Inkling 2 Mini, a 26B-parameter multimodal Mixture-of-Experts model designed for local setups. The model features a novel looped transformer architecture where 12 middle layers process each token three times. It supports a 1M-token context window and accepts both images and text inputs.
The release includes research artifacts that allow users to adjust the number of loop passes to balance speed against quality. Hugging Face provides support for day-zero inference through popular libraries like transformers, llama.cpp, MLX, and SGLang. The company states the model performs competitively against other large models such as Qwen 3.8 27B.
Developers can now run this advanced multimodal model on their own hardware without cloud costs. The ability to tune loop passes offers flexibility for different performance needs. This release expands the options available for running complex AI models directly on local machines.



