Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 show the utility of open models on the Pareto frontier
| Source: Interconnects (Nathan Lambert)
Tags: Inkling, Thinking Machines, Hy3, Tencent, Kimi K3, open weights, MoE, Laguna
The open model ecosystem expanded with Thinking Machines' Inkling (975B-A41B multimodal MoE), Tencent's Hy3 (295B-A21B, now Apache 2.0, reportedly proved a 50-year-old math problem), and Kimi K3—evidence that strong model training is spreading to more organizations than consolidation predictions anticipated.
Details
The latest open model landscape is more crowded than consolidation predictions suggested. Thinking Machines—a company few predicted would become an open-weight leader—released Inkling, a 975B-A41B multimodal mixture-of-experts model supporting text, images, and audio. While not the top performer in its size class among Chinese labs, it is designed as a strong base for fine-tuning through their commercial Tinker platform, which reportedly generates hundreds of millions in revenue annually. Tencent released Hy3, a 295B-A21B MoE with a notable license shift from restrictive terms to Apache 2.0—making commercial use straightforward. The model reportedly proved a 50-year-old math problem using a dedicated harness with Sol as judge, though how much the judge contributed is unclear from the available reporting. Kimi K3 from Moonshot AI uses a revenue-share license rather than fully open source, testing whether partial openness can capture commercial value. The broader trend: more organizations are choosing to release models openly as 'token machines' rather than consolidating—challenging earlier predictions about the market's direction. Note: this is a paid newsletter and not all model details may be present in the excerpt reviewed.