Ten open-weight models dropped between January 27 and February 17, 2026. The original article catalogs all of them in chronological order: Arcee AI's Trinity Large, Moonshot AI's Kimi K2.5, StepFun Step 3.5 Flash, Qwen3-Coder-Next, z.AI's GLM-5, MiniMax M2.5, Nanbeige 4.1 3B, Qwen 3.5, Ant Group's Ling 2.5 and Ring 2.5, and Cohere's Tiny Aya. A March 6 update adds Sarvam 30B and 105B.
The organizing principle is architecture, not benchmark scores. Arcee AI's Trinity Large leads the list: a 400B parameter Mixture-of-Experts model with only 13B active parameters, released alongside Trinity Mini at 26B total and Trinity Nano at 6B total. Each entry is cross-referenced against prior releases, including z.AI's 355B GLM-4.5, to show where design decisions converge and where they diverge. The article explicitly covers QK-Norm, Multi-head Latent Attention, and MoE routing choices across the full set.
The value here is not the summary. It is the side-by-side architecture comparisons and the config-level details that most coverage skips. If you are tracking how Chinese and US labs are converging on similar design patterns, or where they are deliberately diverging, this is the primary source worth reading in full.
[READ ORIGINAL →]