LongCat and Hydrogen launches span multimodal and scientific AI

Meituan has introduced LongCat-2.5, a multimodal mixture-of-experts model trained on Chinese AI chips, and opened API access. It has 1.6T parameters, with approximately 48B active per token, and targets software development, visual design and agent-driven content creation. Deep Principle released Hydrogen-1.0, a 753B open-weight scientific model based on GLM-5.3. Distributed under the MIT license, it is designed for experimental planning, simulation and materials discovery.

Other releases cover voice interaction, robotics and specialized workloads. Rime’s Prism processes speech directly and responds to tone, intonation and interruptions. NC AI introduced a lightweight vision-language model for robots and industrial edge devices, claiming performance above larger competitors. ByteDance Seed published its GenFirst text-to-image repository on Hugging Face. Research releases also include HaloGuard 2.0 for multilingual response safety, Xiaomi Auto World Model 2.0 for driving and embodied AI, and MedPsy for medical language processing on edge devices.

Infrastructure updates emphasize inference efficiency. SGLang v0.5.23 adds performance improvements, broader hardware support and facilities for running models beyond available GPU memory. Zyphra’s hierarchical attention approach uses tree-indexed sparse attention to retrofit existing models. Additional research examines persistent cache failures after brief traffic surges, faster sparse-attention indexing and lower-memory decoding for agent workloads.