The open-model ecosystem is expanding as more companies train strong models, investing hundreds of millions to billions of dollars, and releasing them openly. Thinking Machines, announced in February 2025, now generates hundreds of millions in revenue per year from its open model finetuning service and releases the best open-weight models built in the U.S.A., ahead of NVIDIA with Nemotron and Arcee’s Trilogy. Chinese labs sustain their pace, with newer entrants like Xiaomi gaining mindshare. The article highlights three new releases: Inkling by Thinking Machines, a 975B-A41B multimodal MoE supporting text, images, and audio inputs with text output, positioned as a base for fine-tuning via their commercial offering Tinker, plus a smaller 276B-A12B version. Hy3 by Tencent is a 295B-A21B MoE that improves over its predecessor and switched from a restrictive license to Apache 2; it also proved a 50 year old math problem. Laguna-S-2.1 by Poolside is a newly pre- and post-trained version of the 118B-A8B MoE, marking its third appearance in three consecutive months. The article frames this as an era of continued adoption, questioning how much revenue-share licenses like Kimi K3 can stick and how much market share open models can take.