DeepSeek, Nvidia and Chinese Robotics Firms Report AI Advances

DeepSeek said the new version was trained with self-feedback reinforcement learning and improved its ability to build complete projects from scratch in game development and 3D modeling. It supports more “wish-style” programming tasks and project requests, and token use fell by more than 50% in most scenarios compared with the previous version, according to Kuai Technology and Sina Technology, citing DeepSeek’s official WeChat account. The company also added Search and Fetch interfaces to its API, each priced at 5 yuan per thousand calls.

IT之家 reported that Nvidia’s CentML engineer Benjamin completed an upstream vLLM optimization using large-scale Megakernel fusion and Cute DSL to write Nvidia’s new decode-specific kernel. The Megakernel contains 12,000 lines of code and helped Kimi K3 reach 423 tokens per second per user on GB300 NVL72, 1.68 times the performance on the same hardware, according to the report. Nvidia is also developing vLLM optimizations for DeepSeek V4.1 and other models.

Sanyan Technology reported that Ideal Auto’s chip subsidiary Shanghai Mahexin Technology signed a first-round equity financing agreement worth 1.625 billion yuan. The round was led by Hangzhou Jingtao, Lanchi Ventures and Huaye Tiancheng, with Qiji Chuangtan participating. Hangzhou Jingtao and its affiliated funds received an additional subscription right of up to 500 million yuan; if fully exercised, the post-money valuation would reach 17.1 billion yuan. The funds will support chip R&D and commercialization, especially in non-automotive fields, and the company’s long-term development in chips and domain controllers.

Robot Frontier reported that Hangzhou-based embodied-brain company Gewu Tongzhi completed an angel round of nearly 100 million yuan. Investors included Hongfu Asset, Pike Investment, Hangzhou Capital, Hangzhou Talent Group, a provincial embodied-AI industry investor and several listed companies. The funds will mainly be used for large-scale pretraining and technology iteration of its physical inversion world model, hardware and body construction, talent recruitment, scenario PoC validation and real data collection. Founded in 2026, the company develops self-evolving dexterous manipulation intelligence for real industrial scenarios through a “world model + PSI physical self-evolution” system and says its physical inversion world model lets robots infer contact and force from visible actions for pretraining.

According to a Leiphone digest, JD said it will create more than 100,000 robot maintenance engineer jobs over the next five years. The same digest said OpenAI will develop humanoid robots, with Altman saying the core is to build the robot’s “brain.”