Thinking Machines Releases Inkling 2 Small Trained on 12T Tokens
Thinking Machines Lab has introduced Inkling 2 Small, a Mixture-of-Experts model trained rapidly in 68 hours on 12 trillion tokens. Additionally, METR published a study examining AI research practices and R&D acceleration at Anthropic. This release demonstrates significant efficiency gains in model training times and token scaling for advanced language models. The insights into Anthropic's research methodologies also shed light on how AI tools are accelerating real-world AI development. Inkling 2 Small features 594 billion total parameters with 17 billion active parameters, supports image inputs and variable reasoning effort, and achieves a score of 31 on the Artificial Analysis Intelligence Index.
## BACKGROUND
Thinking Machines Lab develops advanced AI architectures, while METR (Model Evaluation and Threat Research) is an organization focused on evaluating advanced AI capabilities and research practices. Mixture-of-Experts is a neural network design where only a subset of parameters is activated per token to improve efficiency.