Research breakthroughs in efficiency & agents

Climaxing: Mistral AI new open-weight MoE (July early access, 'fat but sparse', $400M+ ARR). PAW paradigm (73.78% accuracy beats Qwen3-32B). Tencent Hy3 (295B total, 21B active, Apache 2.0, 256K context). Meituan LongCat-2.0 1.6T MoE with 1M context on domestic ASICs. GigaWorld-1 for robot policy. Ternlight 7MB embedding model (WASM). New ICML 2026 papers: M+Adam, IT-Opt, HumanLM, Verbalized Sampling (mode collapse fix), constrained decoding for agentic LLMs (guarantees successful tool calls), unlearning with asymmetric sources. Also: Anthropic J-Space interpretability breakthrough; Perceptual Flow Matching (4-8 steps); dOPSD for diffusion LLMs; SaMer token merging; InternVLA-A1.5; EVA-Client; Vision Pretraining; LLM-as-a-Verifier; UI-MOPD. Smaller domain-specific models in pharma.