The predicted consolidation of AI labs has not arrived. Instead, more companies are training frontier-class models at costs of 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, largely ignored as an open-model contender, their fine-tuning service Tinker now generates hundreds of millions in annual revenue and their Inkling model, a 975B-A41B multimodal MoE supporting text, image, and audio inputs, is the strongest open-weight model built in the U.S., ahead of NVIDIA Nemotron and Arcee Trilogy. A smaller variant, 276B-A12B, is competitive in its size class.
Two other releases define this moment. Tencent's Hy3 is a 295B-A21B MoE that improves across all benchmarks over its predecessor and, critically, switches from a restrictive custom license to Apache 2. That license shift matters more than the benchmark gains. Hy3 also produced a proof of a 50-year-old math problem using a dedicated harness with Sol as judge, though the weight of that result is still being assessed. On the revenue-model side, Kimi K3 arrives with a revenue-share license structure, a direct test of whether open models can capture commercial value without full permissiveness.
The question the original piece forces you to sit with is not whether open models survive, but how much market share they can take and under what license terms. Revenue-share agreements like Kimi K3's are either the next evolution in open-model monetization or a ceiling on adoption. Xiaomi and other Chinese entrants are still accumulating developer mindshare. Read the full breakdown for the complete model-by-model analysis, the license comparisons, and the argument for why token demand growth makes consolidation a losing bet for the foreseeable future.
[READ ORIGINAL →]