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

In the world of AI model development, consolidation has been a commonly anticipated path for research labs. While training costs are surging yearly, many companies are still investing billions of dollars in these endeavors. Despite this trend, an increasing number of organizations are opting to release their models openly. This shift is driven by the high demand for tokens and the potential of these models to unlock new use cases.

A prominent example is Thinking Machines, which entered the open model space in 2025 and has since generated significant revenue through their finetuning service. Meanwhile, Chinese labs continue to push forward with newer entrants like Xiaomi gaining traction in the AI economy. As consolidation seems less likely, the focus is shifting towards the continued adoption and potential market share of open models.

One such model is Inkling by Thinking Machines, a 975B-A41B multimodal MoE capable of processing text, images, and audio, and producing text. While not the strongest model in its class, Inkling serves as an excellent base for fine-tuning. Another model, Hy3 by Tencent, is a 295B-A21B MoE that surpasses its predecessor in all metrics.

Importantly, Tencent has opened up the licensing for this release, switching from a restrictive custom license to Apache 2.0. Hybridizing the open model market, Poolside's Laguna-S-2.1 is a 118B-A8B MoE that is small enough to fit on a DGX Spark, making it highly accessible. The company's commitment to transparency is evident in their detailed evaluation trajectories.

Lastly, DeepSeek-ai has recently updated their V4 Flash model, which is currently the best performer on the Pareto frontier. However, it remains to be seen if this will hold true once the bigger model is updated.