Artificial intelligence / news brief
Thinking Machines releases 26B open-weight Inkling 2 Mini for local agents
- Event
- Published

Thinking Machines released Inkling 2 Mini, a 26.4B-parameter multimodal mixture-of-experts model whose middle 12 layers run three times per token. The Apache 2.0 weights occupy about 20GB and are supported on 32GB Apple silicon and DGX Spark. Artificial Analysis scored it 20 on its Intelligence Index, while the model card warns about confident errors, reasoning loops and destructive agent actions.
Why it matters
Looped layers trade extra computation for a smaller stored model, opening a new architecture to local use and outside study without proving broad agent reliability.
Sources
- Introducing Inkling 2 Mini (external site)Thinking Machines Lab · primary source ·
- Thinking Machines has released Inkling 2 Mini, an open weights looped transformer (external site)Artificial Analysis · coverage ·
- Inkling 2 Mini - ARC-AGI Results (external site)ARC Prize · coverage ·