A roundup of six notable open-weight model releases from the past week, with architecture notes on each. Highlights include Nanbeige 4.2 3B using looped depth sharing to double compute without extra memory, Laguna S 2.1 (118B sparse MoE with 8B active params and 1M-token context), Motif-3-Beta introducing Grouped Differential Latent Attention, Solar Open 2 (250B-A15B hybrid MoE by Upstage), Antares 1B (Cisco's cybersecurity-focused small model built on IBM Granite), and BTL-3 (a LoRA adapter for Qwen3.6-27B targeting coding agents). All six models are added to the author's LLM Architecture Gallery.
4 Impressions