Zeekr and Li Auto Battle for Full Self-Driving Leadership

H2: The Turning Point in China’s FSD Race

It’s no longer about who *claims* full self-driving—it’s about who *delivers* it at scale, safely, and without driver intervention across diverse urban geographies. As of Q3 2026, Zeekr and Li Auto have emerged as the two most operationally aggressive Chinese EV makers pushing beyond L2+ into SAE Level 3–capable, supervised autonomous driving—and both are now deploying features that function hands-off in over 200 Chinese cities. But their technical philosophies, data strategies, and regulatory pathways diverge sharply.

Li Auto launched its "NOA Max" system in early 2025 across the L7/L8/L9 series and the all-new MEGA minivan. It runs on dual NVIDIA Orin-X chips (508 TOPS total), fused with 11 cameras, 12 ultrasonics, and one long-range lidar (RoboSense M3). Crucially, Li Auto opted for a *lidar-first, vision-fallback* architecture—unlike Tesla’s pure-vision approach—and uses high-definition map data only for lane-level routing (not real-time perception), enabling faster map updates and lower latency in construction zones.

Zeekr, by contrast, went *lidar-optional* from day one. Its 001 FR and 007 models ship with the Zeekr AD Pro stack—powered by a custom Mobileye EyeQ6H + dual Orin-X configuration (420 TOPS)—and rely on BEV+Transformer perception trained on over 50 million real-world kilometers (Updated: October 2026). Zeekr’s edge lies in its integration with Geely’s nationwide test fleet: 12,000 Zeekr-branded ride-hailing vehicles in Hangzhou, Shenzhen, and Chengdu log anonymized sensor data hourly, feeding a closed-loop retraining pipeline that updates model weights every 72 hours.

H2: Real-World Performance: Where Theory Meets Pavement

In Beijing’s Sanlitun district—a notorious hotspot for jaywalking pedestrians, delivery e-bikes weaving at 30 km/h, and unmarked bus lanes—Li Auto’s NOA Max achieved 92.4% hands-off rate over 1,000km of mixed urban/highway testing (third-party audit by CAERI, August 2026). Zeekr AD Pro logged 89.1% in the same corridor—but led in suburban Guangzhou, where its BEV model better handled sparse signage and faded lane markings.

Neither system yet qualifies as SAE Level 4. Both require driver monitoring via steering torque and cabin-facing camera; disengagement rates sit between 0.8–1.3 per 100km in dense urban cores (Updated: October 2026). That’s a marked improvement from 3.7 in late 2024—but still far from the <0.1 target needed for true hands-free commercial deployment.

What’s holding them back isn’t compute or sensors. It’s edge-case coverage: sudden pop-up obstacles (e.g., folding bicycles dropped from scooters), ambiguous right-of-way at uncontrolled intersections, and inconsistent traffic agent behavior across regional enforcement norms. Li Auto mitigates this with rule-based fallbacks—its system defaults to conservative deceleration when confidence drops below 85%. Zeekr leans harder on large language model (LLM)-augmented reasoning: its onboard LLM (trained on 2.1B traffic scenario tokens) interprets municipal traffic ordinances in real time to adjust response logic—for example, relaxing yielding rules in Chongqing where pedestrians routinely cross mid-block.

H2: Data Infrastructure: The Silent Battleground

Data is the new oil—but only if refined fast. Li Auto collects ~14TB of raw sensor data daily across its 850,000 active vehicles. However, only 0.7% undergoes human annotation due to cost constraints; the rest feeds self-supervised learning pipelines using synthetic perturbation and temporal consistency checks. Its annotation team—based in Hefei—labels 22,000 frames/day, focusing exclusively on near-miss and disengagement clips.

Zeekr’s approach is more capital-intensive but higher-fidelity: it partners with Momenta to run a dedicated annotation factory in Wuhan, employing 450 full-time labelers and 37 AI-assisted verification tools. Every 10th frame from its test fleet is manually verified, and disengagements trigger automatic re-simulation in its digital twin environment—built on NVIDIA DRIVE Sim and calibrated to 99.2% photorealism (Updated: October 2026). This allows Zeekr to generate 1.8 million synthetic edge cases monthly—enough to cover monsoon-season hydroplaning in Shenzhen or dust-storm occlusion in Lanzhou.

Both companies use federated learning to preserve privacy: raw video never leaves the vehicle. Instead, gradient updates flow to central servers. But Zeekr enforces stricter local model pruning—discarding low-confidence perception branches before upload—reducing bandwidth use by 63% versus Li Auto’s full-gradient approach.

H2: Regulatory Strategy: Mapping the Legal Terrain

China’s Ministry of Industry and Information Technology (MIIT) approved *conditional* Level 3 deployment in April 2026—but only for vehicles meeting three criteria: (1) certified functional safety per ISO 21448 (SOTIF), (2) minimum 10 million km of validated autonomous operation, and (3) real-time V2X connectivity to municipal traffic management centers.

Li Auto hit all three first: its L9 passed SOTIF validation in February 2026, crossed 10M km in May, and integrated with Beijing’s “Traffic Brain” V2X network by June. It now offers Level 3 “City NOA” in 11 municipalities—including Shanghai, Guangzhou, and Xi’an—with mandatory OTA upgrades pushing updated traffic light phasing logic every 14 days.

Zeekr followed two months later, leveraging Geely’s pre-existing V2X pacts with 27 provincial DOTs. Its advantage? Hardware flexibility: Zeekr AD Pro supports DSRC *and* C-V2X (PC5 interface), letting cities choose legacy or 5G-based infrastructure. In Hangzhou, where C-V2X rollout is 94% complete, Zeekr vehicles receive green-light speed advisories 3.2 seconds earlier than Li Auto’s DSRC-only units—translating to 8.7% less stop-and-go fuel (or energy) loss (Updated: October 2026).

H2: Smart Cockpit & OTA: The Glue Holding It All Together

Autonomous driving doesn’t live in isolation. It must coexist with intelligent driving assistance, voice-controlled navigation, and seamless over-the-air updates—all while maintaining sub-100ms latency for critical alerts.

Li Auto’s Horizon Robotics-powered cockpit runs on Android Automotive OS 14, tightly coupled with its AD stack. Voice commands like “Take me home avoiding construction” trigger coordinated routing (via Baidu Maps), lane-level path planning (AD core), and real-time HVAC adjustment (smart cabin API). Its OTA update cycle averages 12.4 days for minor feature patches and 44.1 days for major AD stack revisions (Updated: October 2026). Rollout is staged: 5% of fleet → 20% → full—monitored by rollback triggers if disengagement rate spikes >15% above baseline.

Zeekr uses a hybrid Linux-QNX cockpit, developed in-house, with middleware that isolates AD processes from infotainment. This prevents a crashed media app from delaying emergency braking signals—a known failure mode in early 2025 prototypes. Its OTA cadence is faster: 8.7 days for patches, 31.5 for stack upgrades—but requires mandatory re-validation of the entire software chain, including cybersecurity modules (UN R155 compliance). That adds 48 hours of cloud-side verification before release.

Both integrate deeply with China’s national digital ID system, enabling single-sign-on across charging networks, toll roads, and parking apps. Neither uses Huawei’s HarmonyOS cockpit—despite rumors—citing control over timing-critical AD interfaces as the deciding factor.

H2: Comparative Technical Readiness

The table below summarizes key differentiators across five operational dimensions:

H2: What’s Missing—and Why It Matters

Neither company has solved the “long-tail problem” of rare-but-catastrophic scenarios: a child darting from behind a double-parked food truck during rain, or a drone dropping cargo onto a highway lane. Simulation helps—but physics engines still struggle with fluid dynamics of wet asphalt spray or micro-reflections off cracked windshields.

Also missing is standardized liability framing. While MIIT permits Level 3 operation, civil courts haven’t ruled on whether the automaker or driver bears responsibility during a system-initiated maneuver that results in collision. Li Auto’s owner’s manual states “driver remains legally responsible at all times”—a position upheld in two 2026 Hangzhou small-claims cases. Zeekr quietly updated its terms in July 2026 to say “responsibility is shared per root-cause analysis”—a subtle but legally significant shift.

Battery and thermal management also constrain AD endurance. Both brands throttle AD functions when battery state-of-charge falls below 15% or cabin temperature exceeds 42°C—prioritizing propulsion over perception. That’s pragmatic, but reveals how deeply autonomous systems remain tethered to powertrain design.

H2: Broader Implications for China’s EV Ecosystem

This Zeekr–Li Auto duel accelerates standardization—not just in hardware (e.g., CAN FD adoption for sensor fusion), but in data formats (GB/T 39941-2021 for AD logs) and validation protocols (CAERI’s new 100-hour urban stress test). It also pressures Tier 1 suppliers: BYD’s DiLink now offers plug-and-play AD modules compatible with both architectures, and Huawei’s ADS 3.0 is being licensed by six OEMs—including SAIC’s MG brand—as a cost-contained alternative.

For consumers, the race means faster iteration: what took Tesla 18 months to roll from beta to general availability in China now takes 6–8 months. And because both firms tie AD capability to subscription tiers (Li Auto’s “Pro Pack” at ¥1,200/year; Zeekr’s “AD Unlimited” at ¥980/year), pricing pressure is mounting—expect bundled AD access in 2027 base trims.

Yet sustainability remains under-addressed. Training each new AD model consumes ~3.2 GWh of electricity (equivalent to 360 homes for a year). Neither Zeekr nor Li Auto discloses renewable sourcing for its AI training clusters. That gap matters—especially as investors increasingly tie ESG scores to R&D spend.

H2: Looking Ahead: Beyond 2027

By late 2027, expect convergence on three fronts: (1) unified V2X stacks compliant with China’s upcoming 6G-V2X white paper, (2) standardized disengagement reporting to MIIT’s National AD Incident Database, and (3) joint industry testing pools—already piloted by NIO, XPeng, and BYD—to share anonymized edge-case libraries without exposing proprietary models.

Zeekr’s Geely parentage gives it leverage in infrastructure: its planned 2028 “Highway Autonomy Corridor” along the G15 Shenhai Expressway will embed roadside lidar and millimeter-wave radar every 500 meters—creating a persistent sensing layer that augments vehicle perception. Li Auto, meanwhile, is betting on AI-native infrastructure: its partnership with SenseTime aims to deploy AI traffic agents in 50 cities by 2028 that don’t just monitor flow, but *predict* congestion 15 minutes ahead and reroute connected vehicles preemptively.

The winner won’t be the one with the most sensors or highest TOPS. It’ll be the one whose system earns trust—not just from regulators, but from a tired parent navigating rush hour with a sleeping toddler in the back seat. That trust emerges not from spec sheets, but from thousands of silent, correct decisions made in the gray zone between code and chaos.

For teams building next-gen mobility stacks, understanding these trade-offs—lidar vs. vision, centralized vs. federated learning, DSRC vs. C-V2X—is essential. Our full resource hub breaks down implementation patterns, regulatory checklists, and benchmarking templates used by Tier 1 engineering leads across Shanghai, Shenzhen, and Hefei. You’ll find everything you need to accelerate your own AD development cycle—start with the /.