Overview
- Zhipu unveiled GLM-5 with publicly available checkpoints on GitHub and Hugging Face, highlighting stronger coding performance and long-running agent tasks.
- Zhipu reports 744 billion parameters, 28.5 trillion training tokens and a DeepSeek Sparse Attention architecture, with internal tests claiming gains over Gemini 3 Pro on some coding tasks but trailing Anthropic’s Claude.
- Zhipu’s Hong Kong-listed shares jumped about 30% after the launch as Chinese AI stocks rallied on a week of model releases and upgrades.
- MiniMax introduced its M2.5 model, citing in-house benchmarks on coding and search and promoting low-cost usage at roughly 100 tokens per second for an hour at about US$1.
- OpenAI told U.S. lawmakers that DeepSeek employees used obfuscated routes to collect outputs from U.S. models for distillation, with media reports noting ongoing U.S. probes into potential export-control circumvention and records indicating Nvidia provided technical support to DeepSeek.