
Inkling 2 Mini
Visit Inkling 2 Mini websiteMultimodal Mixture-of-Experts model with looped transformer architecture for efficient local AI
- Made by
- Thinking Machines Lab
- Listed by
- Jin
- Pricing
- Open source
- Added
- 20 Nov 2026
- Our coverage
- Read our story
About Inkling 2 Mini
Inkling 2 Mini is a new multimodal Mixture-of-Experts model from Thinking Machines Lab designed for local AI setups. It features a unique looped transformer architecture where 12 middle layers run three times, effectively giving 48 logical layer visits per token. The model has ~26.4B physical parameters with 3.4B active per token (5.4B effective), routes to 8 experts from 128 per layer, supports 1M-token context, and accepts both image and text inputs. Available in NVFP4 and MXFP8 formats with DSpark drafter. Supports Transformers, llama.cpp, SGLang, and MLX. Released under Apache 2.0 license.