Tsinghua University

Three months after founding, SiMate released Simate-beta, a general physical manipulation system using 4D physical perception and hierarchical temporal memory for zero- and few-shot tasks. The company says it tops the RoboDojo leaderboard with an average score of 33.95 and a 27.96% success rate. Internal testers include researchers from MIT, California Institute of Technology, Tsinghua University and Peking University.

SiMate also disclosed an AutoResearch automation platform and Sinfra infrastructure, and said it raised several consecutive rounds in the hundreds of millions of yuan (tens of millions of dollars). Its founders previously shipped a one-stage end-to-end autonomous driving model to mass production. The company plans phased open-sourcing of the model and research tooling by year-end.

Guangxiang Technology and Tsinghua University's Li Shengbo research group have released Phi-WM 1.0 ActEffect, a first-generation world model for physical-native intelligence (Phi), and Phi-bot X1, an industrial self-evolving embodied AI robot aimed at automotive welding loading/unloading and mobile inspection. The robot has completed product validation and entered real-machine deployment on production lines at more than one luxury car plant. The company says deployment is over 10 times faster than traditional industrial automation, post-training real-machine data takes tens to dozens of hours and under 100 hours, and mobile inspection runs over 20% faster than existing line cadence without pre-installed physical infrastructure or fixed cameras.

On the RoboColiseum benchmark built by Zhiyuan Robotics with universities and the open-source community, Yuanli Lingji's embodied foundation model DM0.5 ranks first on all four sub-leaderboards: instruction following (0.8444), spatial understanding (0.6146), disturbance adaptation (0.7344), and general manipulation (0.6370). The company says DM0.5 is currently the only model topping all four boards; it attributes the result to a strong pretraining base plus standard supervised fine-tuning, without task-specific optimization.

The model's core latency dropped from 534 ms to 57.49 ms, and it has native 60-second memory. In logistics sorting it averages about 3 seconds per item with accuracy above 99%, according to the company. Separately, GeoVLA, a collaboration between Yuanli Lingji, Tsinghua University and Tianjin University, was nominated for the IROS 2026 Cognitive Robotics Best Paper Award.

MetaRSI-v1, a meta recursive self-improvement architecture, has been released by CosmosMind jointly with Tsinghua University, Peking University, Stanford University, UC Berkeley, MIT and other institutions. It unifies what the team calls model-level, data-level and harness-level recursion, aiming to improve the process of self-improvement rather than a single model.

With no external teacher model, a 3B active-parameter model (Qwen3.5-35B-A3B) averaged a 10.9-point improvement across Terminal-Bench 2.1, SWE-bench Pro, a GPQA-Diamond subset and AIME; six frontier flagship models averaged a 7.3-point gain on Terminal-Bench 2.1. The RSI-Harness is open-sourced, and the collaboration plans to extend the loop to programmable instruments and automated labs.

Xingce Future has closed A+ and A++ rounds totalling hundreds of millions of yuan (tens of millions of dollars), with investors including Beichuangtou, Meridian Capital China, Xinshang Capital and Xinding Capital. Existing investors Xichuangtou and Houtian Capital joined. Tsinghua University spinoff from the Tian Ge Project team says it has 37 advanced-process GPUs in orbit and 39 payloads running stably, including the first domestic 12nm and global 7nm GPU in-orbit applications. It announced the Fuyao Plan for a cloud-coordinated space-computing constellation and has joined national remote-sensing constellation projects.

At the 2026 China Computing Power Conference, China Mobile Cloud, CETC Nanhu Research Institute, Lynxi Technologies, Iluvatar CoreX, Tsinghua University and Peking University released what they describe as China’s first domestic GPU + neuromorphic chip heterogeneous hybrid LLM inference system. The system splits large-model computation: attention work goes to domestic GPUs while latency-sensitive FFN/MoE expert modules run on neuromorphic chips, coordinated by a self-developed compiler, interconnect protocol and unified inference engine.

In tests with DeepSeek V4, the partners say the setup improves cost-performance more than twofold compared with similar domestic GPU clusters and cuts operating costs by over 40%. It is aimed at token factories, AI code generation, multi-agent collaboration and smart manufacturing.

At the 2026 Inclusion Bund Conference on September 9, Taichu Yuanqi presented its domestic AI-for-science (AI4S) computing platform, which went live in July. The company says the platform runs on its heterogeneous many-core AI chip and is compatible with Loongson, Sunway, Phytium, and x86 CPUs as well as PyTorch, JAX, PaddlePaddle, and MindSpore frameworks.

The company also debuted TecoWeatherNext, a typhoon-tracking system it says fuses meteorological data with AI forecast models to display layered typhoon track, wind field, pressure, temperature, and precipitation. Taichu Yuanqi joined Tsinghua University, Hunan University, Shandong University, Baidu, Dongrun, and nearly 20 other institutions in an initiative for a domestic AI4S computing innovation ecosystem.

Jiyuan Lvdong has released NeoHorse-1, its first agent-native model, in 4B and 9B parameter versions. The company says the model applies multi-model execution experience from its Routing Harness system to agent post-training through route-guided curriculum learning.

Infinigence AI provided compute and infrastructure optimization, while Tsinghua University and Peking University contributed to algorithm research. On 10 agent benchmarks, the 4B version achieves state-of-the-art results among models of the same size, according to the release.

A physics-native world model called Phi-WM 1.0 ActEffect has been released by Guangxiang Technology and Tsinghua University. During training, a controlled world model evaluates action consequences and feeds that signal back into policy optimization; at execution it is removed from the deployment path to reduce latency and compute cost.

On LIBERO, LIBERO-PLUS, and RoboCasa-GR1, the developers report average success rates of 98.8%, 80.3%, and 67.5%. Commercial cooperation in automotive manufacturing is already under way.

Noiz AI, with The Hong Kong University of Science and Technology, Tsinghua University, Carnegie Mellon University, and Google DeepMind, has released HelixWorld 1.0, a real-time interactive audio-visual world model. It generates 24 FPS video and 48 kHz stereo audio together through a native Transformer, supporting interactive world exploration rather than fixed clips. Model weights and code are to be fully open-sourced in the coming weeks, according to the release.