A mid-2026 landscape review of American open-weights large language models, comparing them architecturally and by benchmark performance against Chinese and European counterparts. Key findings: US models are architecturally diverse but lack consensus techniques like Multi-head Latent Attention (MLA) and reasoning-in-pretraining that Chinese labs (DeepSeek, Qwen, Kimi) have converged on. NVIDIA's Nemotron 3 Ultra 550B tops US open benchmarks, while Ai2's OLMo remains the most fully open model globally. Chinese models lead composite leaderboards and global downloads. European labs focus on sovereignty and multilingual coverage. The analysis argues the biggest gap is organizational: most US open models are side outputs of companies whose primary product is proprietary, and there is room for an open-source-first US lab building at 30B+ scale with full training stack transparency.
