Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 show the utility of open models on the Pareto frontier

Summary

The article focuses on the rise of open-weight AI models such as Laguna S2.1, Inkling, and Kimi K3, showing competitive performance at the Pareto frontier despite expectations of industry consolidation.

Why it matters

This shift signals the democratization of AI capabilities; it allows more labs to train and release strong models while maintaining high token demand. It debunks the narrative that rising training costs will force consolidation, instead showing a vibrant ecosystem of open models and commercial fine-tuning services.

Highlights

• Inkling, developed by Thinking Machines: a 975B multimodal MoE positioned as an excellent fine-tuning foundation and offering a competitive smaller version.
• Tencent's Hy3 model: a 295B MoE that switched to an Apache 2.0 license to increase accessibility and solved a 50-year-old math problem.
• Kimi-K3, developed by Moonshot AI: the largest open model in recent times; released under a non-commercial license that requires commercial agreements for inference and fine-tuning.

Critical take

While the proliferation of open models is impressive, the article overstates their sustainability; revenue-sharing licensing and reliance on commercial contracts can limit true openness and hinder broader community adoption.