· 12 min read
Edition 072 — Meituan ships a 1.6-trillion-parameter model trained on Huawei chips, a Berkeley paper trains a robot motion tracker in 3 hours on one GPU, and an 80-million-parameter voice fits inside llama.cpp
Meituan releases LongCat-2.5, a 1.6-trillion-parameter mixture of experts trained entirely on Chinese silicon. Gemini 4 Carbon takes the MathArena crown at 15 times GPT-6.1 Sol's cost, and Claude Fable 5.2 paints with 700,000 simulated brushstrokes. SuperMimic cuts humanoid motion-tracking training from 7 days on 128 GPUs to 3 hours on one. Plus Seedance 2.5, BirdDreams, G-STAR, Soprano in llama.cpp, VibeVoice-ASR, pplx-decider and more.