Spring for Open LLMs: 10 Architectures in January–February 2026

The author presents an overview of ten open LLMs released in January-February 2026, focusing on architectural details. Arcee AI released Trinity Large (400B MoE, 13B active) with local-global attention (3:1), QK-Norm, NoPE, gated attention, and four RMSNorm layers. Moonshot AI introduced Kimi K2.5 (1 trillion parameters) — a multimodal extension of K2 trained on 15 trillion visual and text tokens with early fusion. Step 3.5 Flash from StepFun (196B MoE, 11B active) achieved 100 tokens per second at a context length of 128k using MTP-3. Qwen3-Coder-Next (80B, 3B active) outperformed larger models in coding tasks due to a hybrid of Gated DeltaNet and Gated Attention (3:1). Also mentioned are GLM-5 from z.AI, MiniMax M2.5, Nanbeige 4.1 3B, Qwen 3.5, Ling 2.5 and Ring 2.5 from Ant Group, Tiny Aya from Cohere, as well as the Sarvam 30B and 105B models.