Gemini 3.1 Ultra

A developer tested RAG, fine-tuning, and long-context approaches on the same 1.1M-token knowledge base, finding that long-context cost 24x more than RAG and failed on mid-document facts, while fine-tu…

Z.ai shipped GLM-5.2 on June 13, 2026, with a 1-million-token context window and zero published benchmark numbers. The model, based on a 744-billion-parameter MoE architecture, introduces a dual think…