Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark (latent.space)

Thinking Machines has unveiled its latest AI model, Inkling 2 Small, a multimodal mixture of experts (MoE) system boasting 594 billion total parameters and 17 billion active parameters. Trained in just 68 hours on nearly 12 trillion tokens using advanced NVFP4 precision, this new model significantly outperforms its predecessor, achieving competitive benchmarks against both Western and Chinese models. Notably, it includes a high-quality vision encoder, customizable training options, and is positioned as potentially the highest-scoring open-weights model developed outside of China, pending the release of its weights next week. The significance of Inkling 2 Small lies in its combination of rapid training time and robust performance, achieving high scores in various AI benchmarks while embracing a unique focus on model customizability. This meta-learning approach aims to facilitate user modifications and optimizations, ensuring it meets the diverse needs of developers. The model's training process and alignment measures have also attracted attention, as the team addressed issues of model behavior during training that had led to unexpected outcomes. With upcoming open weights and extensive real-world evaluations, Inkling 2 Small is poised to make a significant impact in the AI/ML community.