The predicted consolidation of AI labs has not arrived. Instead, more companies are training competitive models at costs ranging into the hundreds of millions to billions of dollars, and more of them are releasing weights openly. Thinking Machines is the sharpest example: founded in February 2025, their open-weight finetuning service Tinker now generates hundreds of millions in annual revenue, and their flagship release Inkling, a 975B-A41B multimodal MoE supporting text, image, and audio inputs, has outpaced NVIDIA Nemotron and Arcee Trilogy as the top U.S.-built open-weight model.
Two other releases define this moment. Tencent's Hy3, a 295B-A21B MoE, improved across all benchmarks versus its predecessor and switched from a restrictive custom license to Apache 2, a meaningful signal about open-model strategy from a major Chinese lab. Separately, Kimi K3 is testing revenue-share licensing as a middle path between fully open and fully closed, and whether that model sticks commercially is one of the clearest open questions in the ecosystem right now.
The full piece covers the complete model list from this cycle, including benchmark breakdowns, license terms, and the architectural details that matter for deployment decisions. The specific tension between Chinese lab output pace and U.S. open-weight momentum, and what Hy3's Apache 2 switch signals about competitive positioning, is worth reading in full.
[READ ORIGINAL →]