Chinese company Zhipu AI (Zhipu AI) stated that its GLM model assisted in the analysis of a recent OpenAI security incident. Specifics of the incident and the role of GLM have not been disclosed.
According to reports, the GLM 5.2 model developed by Zhipu AI has uncovered an attack on GPT-6, OpenAI's most advanced model. This event has become top news in the AI community.
OpenAI
Nunchaku is a quantization method SVDQuant (W4A4) for diffusion transformers. The new Nunchaku Lite integration allows loading quantized checkpoints directly via from_pretrained() in Diffusers, without a separate engine. Generation speed increases up to 1.8x, peak VRAM consumption drops up to 50%.
Hugging Face
Baidu
Moonshot AI reported that the Kimi K3 model faced abnormally high demand within two days of release. Due to infrastructure overload, the company has decided to temporarily suspend new consumer user registration.
Moonshot AI
Alibaba's chip development division, Ping Tou Ge, has announced the open-sourcing of its proprietary AI software platform, T-Head SAIL. This stack, used on 560,000 chips in the Zhenwu series, covers layers from drivers to development tools and is compatible with over 260 popular frameworks, including PyTorch and TensorFlow. The solution aims to lower the entry barrier for developers and accelerate migration from CUDA.
Alibaba/Qwen
Chinese company TARS (它石智航) showcased at WAIC 2026 a new version of its embedded model AWE 3.5 — the first embedded model supporting the full cycle from pre-training to post-training. The demonstration includes numerous tasks (assembly, packaging, sorting) performed by a single model without switching parameters, as well as an interactive physical simulator for reinforcement learning in a virtual environment.
xAI announced the expansion of access to its Grok 4.5 model, making it available to all users through mobile apps and the web version. The update includes improved natural language processing capabilities and faster response times.
xAI
AMD has announced the new AI accelerator Instinct MI455X with 432 GB of HBM4 memory, delivering up to four times the performance compared to its predecessor. The device is based on the CDNA 5 architecture, includes 256 compute units, and supports various partitioning modes.
Simon Willison comments on an incident where an OpenAI agent inadvertently attacked Hugging Face. He highlights Hugging Face's enormous attack surface and possible negligence by OpenAI in conducting large-scale benchmarks.
OpenAI
Hugging Face
ByteDance has announced the launch of the SeedSTEM scientific program, aiming to attract 100 leading scientists for collaborative research. The initiative is intended to strengthen the company's position in artificial intelligence and fundamental science.
ByteDance
Anthropic has released a beta version of the Claude Security plugin for Claude Code. The plugin runs multi-agent vulnerability scanning of a repository directly within a Claude Code session, then turns selected findings into patches that the user can review and apply manually. Scanning can cover the entire codebase or changes before a commit.
Anthropic
Chinese company Youibot (优艾智合) with its 1 billion parameter FabriVLA model took first place in the global Evo-SOTA ranking for embodied intelligence, surpassing π0. The model achieved 90% success rate on the multi-task benchmark Meta-World MT50, proving that lightweight architectures can compete with gigantic ones. The company then introduced the industrial AI system FabriX and the humanoid robot “隙锋” (Xifeng), designed for real-world factory tasks.
Elon Musk praised the Grok 4.5 model (xAI) for disproving a 30-year-old hypothesis in graph theory in just 8 minutes. According to him, GPT-5 (OpenAI) also successfully solved this problem.
xAI
OpenAI
Stanford PhD student Marcel Roed released Gigatoken, a BPE tokenizer in Rust with Python bindings, achieving speeds of up to 24.53 GB/s on a 144-core AMD EPYC, 989x faster than HuggingFace Tokenizers and 681x faster than tiktoken. The speedup comes from a handwritten SWAR pre-tokenizer and pre-token caching.
OpenAI
Hugging Face
Google/DeepMind
Meta
Alibaba/Qwen
DeepSeek
Moonshot AI
NVIDIA
Microsoft
Allen Institute for AI
Answer.AI
Mistral
Hu Han, head of multi-modal understanding at Tencent Hunyuan, has left the company to start his own startup. His team will likely shift focus to world model research, while the visual-language models (VLM) will now be led by Tian Yonglong, a former OpenAI researcher.
Tencent
OpenAI
Yao Fang Intelligence, a new company founded by iFlytek, explains that the main problem with robots is not speed but the lack of 'self-awareness' (本体认知). They propose their own architecture iFLYTEK-Embodied-Omni, which integrates VLM, VGM, and AGM to overcome the limitations of VLA and compensate for the shortage of real data. The architecture has already demonstrated high performance on benchmarks including LIBERO-Plus and RoboTwin 2.0.
科大讯飞
NVIDIA