News
November 21, 2026 · Hugging Face

Thinking Machines releases Inkling 2 Mini, an open looped model for local AI

Inkling 2 Mini is an Apache 2.0 multimodal mixture-of-experts model with about 3B active parameters, looped layers and a 1M-token context, with llama.cpp and MLX support.

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, an open-weight multimodal model aimed at people who run AI on their own hardware. The release was detailed in a Hugging Face blog post and model card, which list the licence as Apache 2.0.

What's new

  • Small active footprint: Hugging Face describes a mixture-of-experts model with about 26B total parameters, of which roughly 3B are active for each token.
  • Looped layers: the model has 24 physical decoder layers, but the 12 middle ones are run three times using the same weights, so each token passes through 48 layer visits. The blog says this is a way to gain capacity without storing more distinct layers.
  • Inputs and context: it accepts text and images and supports a context window of up to 1M tokens, according to the post.
  • Shipping format: expert weights come in mixed NVFP4 and MXFP8 precision, and a drafter model for speculative decoding is included to speed up generation.
  • Runtime support: Hugging Face lists support in Transformers, llama.cpp through its Llama App, Apple's MLX and, for data-centre deployments, SGLang from day one.
  • Research variant: a separate checkpoint can run one, two or three loop passes. Fewer passes should be faster but may lower quality, and the blog says it has not been evaluated in depth.
  • Benchmarks: the blog reports that Inkling 2 Mini leads in 32 of 48 benchmarks when set against Qwen3.8-27B, Nemotron 3.5 Super VL, Muse Glimmer and Gemma 4 31B. These are the publishers' own figures.

Why it matters

Analysis: a model with only a few billion active parameters, a permissive licence and ready-made builds for llama.cpp and MLX is the kind of release that local AI users can try the same day. The looped design is the interesting part: it trades extra compute per token for a smaller file, which suits laptops and desktops where memory is the tighter limit. Treat the benchmark lead with care until independent tests arrive, since the comparison comes from the release partners themselves.

Source: Hugging Face