Consolidation in AI labs was supposed to be inevitable. Training costs rising by orders of magnitude per year, the logic went, would force mergers by 2026 or 2027. That prediction is wrong. More companies are training strong models, spending hundreds of millions to billions of dollars, and more of them are releasing weights openly. Thinking Machines launched in February 2025 and nobody called them an open-model company. Now their fine-tuning service Tinker generates hundreds of millions in annual revenue and their flagship Inkling, 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.
Two other releases define this moment. Tencent's Hy3 is a 295B-A21B MoE that improves on its predecessor across all benchmarks and, critically, drops its restrictive custom license for Apache 2. That licensing shift matters more than the benchmark gains. Hy3 also used a dedicated harness with Sol as judge to prove a 50-year-old math problem, though the full significance of that setup is still being scrutinized. On the Chinese side, newer entrants like Xiaomi continue accumulating developer mindshare while Kimi K3 tests whether revenue-share licenses can hold at scale.
The original piece is worth reading in full because the model list is extensive and the framing question is the right one: not whether open models survive, but how much market share they can take in the decisive period now beginning. The spec breakdown on Inkling's smaller 276B-A12B variant and the detail on Hy3's benchmark-by-benchmark improvements against its predecessor are both in the full recap.
[READ ORIGINAL →]