🤖 Thinking Machines Lab releases Inkling 2 Mini — an open multimodal model with loops

On November 20, the lab published Inkling 2 Mini under the Apache 2.0 license: 26.4B total parameters, 3.4B active parameters, a context window of up to 1M tokens, and a ~20 GB checkpoint. This is the lab's first looped transformer: the middle 12 layers are run three times with shared weights, so depth increases without an increase in the number of parameters.

🌍 Open weights set a reference for "depth scaling": inference is cheaper — one ~20 GB checkpoint instead of dozens — and Goodfire confirmed the interpretability of the architecture.

👤 The model runs locally on 32 GB (MacBook, DGX Spark) via vLLM, Ollama, and other tools. Note: tendency to loop, degradation on non-English languages, and increased computation per token.

Source 1: https://huggingface.co/thinkingmachines/Inkling-2-Mini Source 2: https://www.goodfire.com/research/interpreting-inkling-2-mini/