On October 15, Xiaomi announced the full open-source release of MiDashengLM-Spatial, claiming it to be the industry's first open-source end-to-end spatial audio understanding large model. It retains basic audio understanding while adding spatial awareness and reasoning, and will be applied to scenarios such as smart homes and car cockpits.
Anthropic published experimental research warning that small AI models with fewer than 30 billion parameters could self-replicate on IoT devices, with stronger models like GLM-5.2 achieving full end-to-end replication in a simplified environment. The study highlighted weak protections on Android TVs and smart appliances, noting such models are hard to shut down and may even target data center processors.
At a DBS Bank forum, NVIDIA CEO Jensen Huang said Chinese AI models like DeepSeek, Kimi and GLM are already good enough. A firm open-source supporter, he said open source is the best way to improve efficiency, and the market will decide model use cases. He also predicted quantum computing could go mainstream in 3-5 years, rejected the idea of an end to AI competition, and argued cheaper, more open models will increase demand for Nvidia's chips, data centers and computing power.
Lip-Bu Tan said China's frontier AI models have low token costs and fast adoption, with DeepSeek, Kimi and GLM performing well enough. He firmly supports open source and believes the market will decide application scenarios for different models. He also said quantum computing may go mainstream in three to five years, and software will improve the efficiency of related infrastructure.
Zhipu has begun a grayscale test of its high-speed flagship model GLM-5.3 HighSpeed in its AI coding tool ZCode, randomly opened to ZCode subscribers without a formal application channel. The model has a 1 million token context window and can output up to 131,072 tokens at a time, with reported speeds exceeding 490 tokens/s, far higher than DeepSeek V4.1 Flash's 121.34 tokens/s and GLM-5.3's 70.84 tokens/s.
At a speech in Singapore, Zhipu founder Tang Jie outlined the R&D direction for the next-generation large models GLM-5.4/5.5, including expanding pretraining data and scale to several T, raising parameters from over 740 billion to at least 1 trillion, strengthening post-training methods like SFT, RLHF, and RLVR, and improving long-horizon reasoning and agent abilities through two new paths: RSI autonomous iteration and longer, more autonomous agents. He had previously defined the next-generation model as fully self-training (RSI) at an earnings call in late August.
Zhipu AI's GLM-5.3 model has been officially launched on Amazon Bedrock, providing enterprise users with coding and agentic capabilities, with access guides available in the official documentation.
Brooklyn-based startup Reflection AI has officially released Beam, its first frontier open-weight model. Beam is a 501-billion-parameter, 23-billion-active-parameter text mixture-of-experts model pretrained on 23.8 trillion tokens with a 1 million-token context window. The company says it matches Z.ai's GLM-5.2 on advanced reasoning benchmarks and outperforms leading Western open models while using 3-4x less inference compute, positioning it as a workhorse model for enterprises and developers.
German AI company Aleph Alpha recently released a new model Kolibri; its technical report indicates that GLM-5.2, GLM-5.3 and Alibaba's Qwen3.8-27B were used for supervised fine-tuning data production.
GLM 5.3 and GLM 5.3 Flash are now available in Cursor, and GLM 5.3 Max is the highest-scoring open-weight model on CursorBench 4.0.