New Open-Weight Models and Research Advances

Google's Gemma 4, Thinking Machines Lab's Inkling (975B/41B active), Kimi K3 (3T mxfp4), and Tencent's Hy3 295B MoE push open-weight frontier. Research includes StateAct for computer-use agents, multimodal dataset alignment best practices, and new RL training frameworks (OpenForgeRL, Molt). New papers: Mage-VL (efficient streaming multimodal), OmniDelta (token compression), VIPE (video prompt engineering), Keep It InMind (agent memory blind spot), PerceptionBench (MLLM perception), Pass the Baton distillation, Parallel Decoding Distillation, and RL training stability improvements. Infrastructure patterns for agentic applications highlight queue-based control flow.