Li Auto held its Livis Day software and embodied intelligence conference on June 15, 2026, with CEO Li Xiang delivering a comprehensive vision of the company's transformation from automaker to embodied intelligence company. Unlike traditional car launches, the event focused entirely on AI technologies.
Mach M100: World's First Dataflow AI Chip
The centerpiece was the Mach M100, the world's first dataflow architecture AI chip on a 5nm automotive-grade process. Each chip delivers 1280 TOPS with over 82% actual utilization efficiency. The dual-chip configuration (2560 TOPS total) enables a 50% increase in imitation learning scale, 15x reinforcement learning, 10x model parameters, and 15x computation.
Li Auto abandoned the von Neumann architecture for a dataflow design where data movement triggers computation. Over half the chip die is dedicated to a neural network processor with 56 compute units connected via grid bus and data ring bus dual-interconnect architecture. The CPU section has 24 Cortex-A78AE cores at 2.3GHz with 8-channel LPDDR5X at 273GB/s.
In benchmarks, dual Mach M100 demonstrated multiple times the performance of NVIDIA Orin across CNN backbones, UniAD, and VLM models. Running Qwen 1.5B and 3B models, a single vehicle-mounted M100 outperformed the $4,000 NVIDIA DGX Spark desktop supercomputer with 2.7x prefill and 1.5x decode speeds.
The architecture paper was accepted by ISCA 2026, making Li Auto the first automaker ever to have a paper accepted at ISCA's industrial track. The team presents on June 30.
Mach Mind-4 Series
Li Auto unveiled two foundation model variants:
-
Mach Mind-Pro: Top-tier across IFEval, LongBench-v2, AIME26, BFCL-v4 benchmarks. In Agent evaluations it outperforms most mainstream models. Compared to a closed-source Model S, it reduces token consumption by 38% and redundant tool calls by 47%, with 208 tokens/s peak TPS. Already deployed in Li L9 Livis.
-
Mach Mind-Edge: The industry's first mass-produced on-device embodied intelligence agent, using multimodal streaming temporal modeling. It performs always-on active perception, interaction, autonomous control, and multimodal Q&A - all locally on the vehicle with zero data uploaded.
Mach VLA Roadmap
- Q3 2026: All-new Mach VLA for AD Max
- Q4 2026: Align with Tesla FSD V14
Reaction time of 0.28 seconds, approaching F1 driver threshold (0.25s vs human 0.45s). At 120km/h, this means ~6 meters less braking distance.
OTA Plan: July, September, December
July: 30% driving efficiency gain, narrow passage assist, travel Agent skills. September: human-like reversing, smart lock/garage, ultra-wide CarPlay. December: safety surpassing humans, police gesture recognition, 0.2s reaction time.
Redefining the Intelligent Vehicle
Li Xiang defined the embodied intelligent vehicle as 'four in one': an electric vehicle, a professional driver, an AI computer, and a life assistant. He argued smartphones and smart cars today are 'function-driven,' not truly intelligent. True embodied intelligence must protect human safety, complete tasks independently, and operate more efficiently than humans.
The company disclosed 17.27 million active collision avoidance interventions by its AD systems, including 55,000 critical ones. It reorganized its foundation model team into three departments: embodied engineering, embodied interaction, and embodied behavior.
Li Auto's next major launch is the Li L8 on June 23, available in Ultra and Livis trims (Livis with dual M100 chips).
