Introducing Inkling 2 Small: A 594B MoE Model with Open Weights Coming
Thinking Machines has announced Inkling 2 Small, a Mixture-of-Experts model featuring 594B total parameters and 17B active parameters, along with multimodal support and variable effort. The model is initially available through Tinker with closed weights, and its full weights will be released under an Apache 2.0 license next week following safety testing. This release represents a significant generation upgrade in open-weight models, achieving notable performance gains in agentic coding, vision, knowledge work, and reasoning. The upcoming open-weights availability under an Apache 2.0 license will provide developers and researchers with powerful tools for fine-tuning and deployment. Inkling 2 Small far surpasses previous versions with a score of 31 on the Artificial Analysis Intelligence Index. Improvements stem from enhanced pre-training and a large, unified reinforcement learning run.
## BACKGROUND
Mixture-of-Experts (MoE) is a machine learning architecture where multiple specialized sub-networks ('experts') are used, but only a subset of them is activated for any given token to maintain high performance with lower compute costs. An Apache 2.0 license is a permissive open-source license that allows users to freely use, modify, and distribute the software or model.