从李汽车展日看：迈入智能体时代

Li Auto June16,2026 Key Takes from Livis Day: A Step into theAgent Era as Al agents by integrating autonomous driving, in-cabin interaction, localagent capabilities, and compute into a unified stack. With 2,560 TOPS ofAD Max compute, 0.28s end-to-end VLA latency, and a 4Q26 target toreach parity with Tesla FSD V14, this increasingly reads as a true verticalintegration story rather than a branding exercise. a "tool that drives"into an'Al agent that gets things done."The idea rests on three points: 1)Safety first: instead of handing control back to the driver at critical moments, the system aimsto handle risks itself and prioritise protecting occupants (management noted that~50% of L2+accidents happen during handover). 2) Beyond driving: the goal is not just better driving, butcompleting end-to-end tasks autonomously. 3) Higher efficiency than humans: the systemshould ultimately drive more safely and efficiently than a human. The company frames theproduct as "EV + professional driver + Al computer + life assistant." Importantly,this sameframework will be appliedto future humanoids and otherembodied Al products,placing theauto business within a broader robotics roadmap. Agent stack: Mach Mind (cloud + on-device). Mach Mind is Li Auto's unified Al stack acrosscloud and vehicle. Mach Mind Pro (cloud) handles language and complex tasks with top-tiercapability, while improving efficiency and latency for in-car use. Mach Mind Edge runs on-device as a multimodal agent, enabling real-time interaction and control without data upload,supporting both low latency and privacy. Together, it combines cloud intelligence with on-deviceresponsiveness,allowingboth complextaskhandlingandreal-timeexecution M100: a real in-house chip story. Li Auto's Mach M100 is a 5nm, 1,280 TOPS Al chip builton a dataflow architecture, enabling >82% utilization, well above typical GPUs.Over half of thechip is dedicatedto neuralnetwork compute, supported by a 24-core CPUand high-bandwidthmemory. Performance stands out, with M100 reportedly outperforming NVIDIA Thor on ADworkloads and exceeding DGx Station in LLM inference,positioning it as a unified computeplatform for both autonomous driving and Alagents. The chip is tightly integrated with Li Auto'sin-house stack and Lios, helping deliver ~0.28s end-to-end latency. Clear OTA roadmap with step-by-step milestones, targeting human-level outperformanceby end-2026 and Tesla parity by 4Q26. Li Auto's smart driving is advancing quickly, with0.28s end-to-end response time already faster than human drivers and equating to ~6mshorter stopping distance at highway speeds. This is backed by 17m+ real-world safetyinterventions and newfeatures such as suspended-obstacle detection and improved nightrapid data expansion, and a shift to a unified multimodal architecture. Mgmt is guiding for~30% efficiency improvement by Jul-26, with a clear path to FSD V14 parity by 4Q26 andhuman-leveloutperformanceby end-2026. Xiaoyi Lei * |Equity Analyst85237671126|xiaoyi.lei@jefferies.com AaronWang*EquityAnalyst+85237671121laaron.wang@jefferies.com Jefferies Li Auto Li Auto Inc.manufactures automobiles. The company designs, develops, and sells smart electric sport utilityvehicles. Li Auto conducts businesses in China. Li Auto is one of the key smart EV makers in China. The company designs, develops, manufactures, and sells premium smart electric SUVs. Li Auto is the first to successfully commercialize extended-range EVs in China. The company's first model, Li One, was introduced in Nov 2019. The CompanyValuation/RisksLi Auto policy towards EREV, and 3) setbacks in autonomous driving development Analyst Certification:I, Xiaoyi Lei, certify that all of the views expressed in this research report accurately reflect my personal views about the subject security(ies) and subject company(ies). I also certify that no part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or viewsexpressed in thisresearch report.I, Aaron Wang, certify that allof the views expressed in this research report accurately reflect my personal views about the subject security(ies) and subject company(ies). I also certify that no part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or viewsexpressed in this research report.Registration of non-US analysts: Xiaoyi Lei is employed by Jefferies Hong Kong Limited, a non-US affliate of Jefferies LLC and is not registered/qualified as a research analyst with FINRA. This analyst(s) may not be an associated person of Jefferies LLC, a FINRA member firm, and therefore may not be subjectto the FINRA Rule 2241 and restrictions on communications with a subject company. public appearances and trading securities held by aresearch analyst.Registration of non-US analysts: Aaron Wang is employed by Jefferies Hong Kong Limited,a non-US affiliate of Jefferies LLC and is not registered/qualified as a r