Li Auto's fundamental models team unveiled an artificial intelligence trilogy on September 22nd, consisting of the ME-Brain-1.0 system, the unified model for understanding and generating actions in the MachEmbodied world (ME-U0), and the MachEmbodied-VLM cognitive models (ME-VLM). The official English name for these developments is Li Auto / MachEmbodied. The main focus of this review is on the architecture of the system and models—memory, cognition, and action—rather than on marketing electric vehicle products.
According to the ME-Brain-1.0 technical blog, this system integrates Evolvable Memory, a Cognitive Core, and an event-driven Action Model into a cycle that includes execution, experience logging, and skill updating. The Cognitive Core is available in an MoE variant with 35B-A3B parameters and in a compact version for edge devices with 4B parameters. In laboratory tests, the large core achieved an average score of approximately 70.9 on embodied systems datasets and 72.5 on agent datasets, while the 4B parameter version ranked second compared to other companies.
The Evolvable Memory and its associated modules operate on Li Auto's peripheral M100 SoC to generate and retrieve memory directly on the device. ME-VLM represents a path of the Cognitive Core, where quantization W4A8 and token compression reduced the prefill latency on the M100 from approximately 400 ms to 188 ms.
The ME-U0 model was pre-trained on approximately 4200 hours of data selected from larger robot datasets and first-person data. According to the company's metrics, the model achieves an average success rate of 99.1% on standard LIBERO, 81.8% on LIBERO-Plus without specific adaptation, and a process score of 17.66 on RoboDojo-Sim among participating world action models. The architecture uses dual-expert modules for understanding and generation with multi-speed rotational encoding of position, ensuring temporal alignment of visual latent representations and action tokens.
External publications, including coverage by Robot Forward republished by Phoenix Tech, note visible capture pauses, periodic capture errors, and incomplete tasks in open scenarios, even where scripted scenes succeed. These observations are useful additions to laboratory leaderboards. For now, the benchmarks remain data provided by Li Auto until external labs reproduce the results. The short-term signal is the existence of a packaged memory-cognition-action stack with a documented path on a peripheral SoC, rather than the presentation of a single chassis.
