Thinking Machines Lab releases Inkling 2 Mini, a looped-transformer MoE multimodal model with open weights
Original title: thinkingmachines/Inkling-2-Mini (MoE: 26b, a3B)
Thinking Machines Lab released Inkling 2 Mini, an open-weights multimodal looped-transformer MoE model with 26.4B total and 3.4B active parameters (about 5.4B effective active per token due to three passes through the looped middle layers). It supports text and image inputs, a 1M-token context window, Apache 2.0 licensing, and runs locally on a single NVIDIA DGX Spark or 32GB+ Apple silicon/GPUs via vLLM, SGLang, llama.cpp, Ollama and other frameworks. Benchmarks show strong coding and agentic results against similar-size models, while the card flags limitations including repetitive reasoning loops, lower reliability in agentic tool use, and weaker non-English output.
The official model card provides architecture details, local hardware requirements, and benchmark comparisons, helping readers assess its fit for local deployment and agentic use.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.