en.Wedoany.com Reported - At the inaugural Livis Day event held on June 15, Li Auto announced that its Mach VLA capabilities will continue to be upgraded in the second half of the year. According to the plan, a new Mach VLA will be pushed to the AD Max intelligent driving system in the third quarter, while alignment with Tesla FSD V14 capabilities will be achieved in the fourth quarter.
Li Auto introduced the hardware foundation behind the Mach VLA. The dual Mach M100 chip delivers 2560 TOPS of computing power, with a 50% increase in imitation learning scale, a 15-fold increase in reinforcement learning scale, a 10-fold increase in model parameter count, and a 15-fold increase in model computation. At the event, Li Auto also unveiled its self-developed new base model series—the Mach Mind-4 series. This series includes two versions: Mach Mind-Pro and Mach Mind-Edge. Among them, the Mach Mind-4 series focuses on language intelligence, while the Mach VLA focuses on machine intelligence.
The Mach Mind-Pro has demonstrated outstanding performance across multiple authoritative benchmarks, including IFEval instruction following, LongBench-v2 ultra-long text comprehension, AIME26 advanced mathematical reasoning, and BFCL-v4 tool invocation. Additionally, in the specialized Agent domain, the Mach Mind-Pro's overall performance is at the industry-leading level on the Claw series benchmarks and the PinchBench-v2 real-world tests. According to Li Auto, compared to a certain closed-source model S, the Mach Mind-4-Pro reduces average token consumption by 38%, decreases redundant tool invocation rounds by 47%, and achieves a peak model TPS of 208 token/s. The core of Mach Mind-Pro is its native Agent intelligence capability, which has been fully implemented in the L9 Livis in-vehicle intelligent system.
The other model, Mach Mind-Edge, is positioned as "the industry's first truly mass-produced and deployed on-device native embodied intelligent agent." It adopts multimodal streaming temporal modeling, enabling continuous understanding of the dynamic physical world, along with causal reasoning and autonomous decision-making capabilities, allowing it to directly output actions and call vehicle hardware in real time. Mach Mind-Edge supports tasks such as all-weather active perception, human-vehicle interaction, autonomous vehicle control, and multimodal Q&A, with all data processing completed locally on the vehicle side and no data uploaded to the cloud. Li Auto emphasized that Mach Mind-Edge is not a stripped-down version of a cloud model, but rather an on-device native embodied intelligent agent built from the ground up for in-vehicle scenarios.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
