Inkling 2 Mini: Thinking Machines Lab Releases Looped Transformer Model for Local AI
Thinking Machines Lab has unveiled Inkling 2 Mini, a new multimodal Mixture-of-Experts model designed for efficient local AI setups, featuring a unique looped transformer architecture.

By Clara Tan, Artificial Intelligence
3 min read
Thinking Machines Lab has introduced Inkling 2 Mini, a new multimodal Mixture-of-Experts (MoE) model engineered for cutting-edge performance on local AI setups. This release follows the Inkling 2 Small and aims to bring advanced AI capabilities to more accessible hardware.
Inkling 2 Mini stands out with its innovative "looped transformer" architecture. While it has 24 physical decoder layers, the middle 12 layers are run three times in a loop, effectively giving it 48 logical layer visits per token. This design allows the model to achieve higher capacity with fewer physical parameters, making it more efficient for local deployment.
The model boasts approximately 26.4 billion physical parameters, with 3.4 billion active parameters per token. Due to its looped architecture, the effective active parameters per token increase to 5.4 billion. At each layer visit, Inkling 2 Mini routes the token to 8 experts selected from 128 experts per layer, a more compact configuration compared to Inkling 2 Small's 384 experts.
Inkling 2 Mini is natively multimodal, accepting both image and text inputs, which are projected into a shared hidden space and processed jointly. It supports an impressive 1-million-token context window, enabling it to handle extensive inputs and generate comprehensive text outputs. The model is shipped in mixed-precision formats, including NVFP4 and MXFP8 expert weights, and includes a DSpark drafter for additional speed-up.
Thinking Machines Lab provides support for various inference frameworks, including standard Transformers, Llama App (llama.cpp), SGLang, and MLX. Initial "vibe tests" suggest its performance is comparable to models like Qwen 3.8 27B, positioning it as a promising option for developers building AI-powered applications, agentic systems, coding assistants, and chatbots.
Inkling 2 Mini is released under the Apache 2.0 licence, with open weights available for download on Hugging Face. It is intended for general-purpose conversational use, instruction-following, and other natural language and visual tasks, supporting English and other languages, as well as multiple coding languages.
Sources
- 1Hugging Face Blog: Welcome Inkling 2 Minihuggingface.co
- 2Thinking Machines Lab: Inkling 2 Mini Model Cardthinkingmachines.ai