Can NIO’s World Model Compete with Tesla and Li Auto?

Since last week, rumors, impressions, and speculations about NIO’s World Model (NWM) have been brewing in public discourse.

For instance, the first “locked version” experience video posted on May 16 by @Harry_AD, NIO’s Head of Intelligent Driving Products and Experience:

Or the “first-tier” impressions circulating today (May 21) in NIO owner and discussion groups:

There’s also the flood of commentary from bloggers on the Chinese social media platform Weibo evaluating the initial performance of the world model:

Opinions vary, but one thing is certain: NIO’s World Model is poised for launch.

The remaining question is: Can it actually compete?

With ADS 4 seizing the spotlight as the “first commercial L3 release” and Li Auto’s VLA roadmap entering its rollout phase, can NIO stage a dramatic comeback this year?

While we’ve signed NDAs, here’s what we can share based on available information.

1. What is NIO’s World Model?

The concept of a “world model” isn’t new to NIO—it’s an established term in AI research dating back decades.

Outside the automotive sphere, the most buzzworthy recent example is Google’s Genie 2, which generates diverse 3D worlds. Users can define worlds using natural language, and Genie 2 even simulates character movements within them.

Similarly, NIO’s NWM allows its autonomous driving system to simulate varying road conditions for the same scenario (e.g., generating left turns, right turns, or straight paths) and produce synthetic videos.

The foundation version of NWM is a “training” model, not the “inference” model deployed in vehicles. Its purpose is to “define the rules of the world,” while the vehicle’s task is to “interpret those rules.”

As William Li mentioned earlier, NWM can be distilled into Loctek’s Horizon Robotics chip solutions. Meanwhile, NT2’s quad-Orin and NT3’s next-gen chips will also receive world model updates.

@Harry_AD’s official statement clarifies:

“Loctek models are expected to receive experience updates based on end-to-end models later this year, ensuring the L60 remains competitive. Future Loctek vehicles will adopt NIO’s world model architecture for ADAS solutions.”

This suggests NIO’s diverse hardware platforms will share a unified technological foundation.

2. Is NIO’s World Model End-to-End?

Theoretically, all generative AI is end-to-end—from input to output.

Since Tesla’s 2019 Highway NOA, automakers and suppliers have spent six years refining how to make ADAS systems more seamless and consistent.

Marketing terms like “nationwide coverage,” “one-stage end-to-end,” or “two-stage end-to-end” are merely milestones in this evolution.

World models, foundation models, end-to-end architectures, and VLA (Vision-Language-Action) frameworks aren’t directly comparable.

For example, Waymo’s EMMA (End-to-End Multimodal Model for Autonomous Driving), released in November 2023, shares technical principles with VLA. Waymo notes EMMA was “developed” using Google’s Gemini foundation model.

3. Can NIO’s World Model Compete?

Technically, world models align with first-principles thinking, providing a unified training framework for ADAS and future autonomous driving that mirrors human intuition.

What intuition? Generalization.

Since the 1971 introduction of “mental models,” cognitive science and AI research have aimed to create systems that perceive, reason, and decide like humans. This led to concepts like counterfactual reasoning, enabling AI to answer “what if” questions—essentially imagining possibilities.

Humans generalize effortlessly; AI requires scaffolding. World models, as seen in Google’s Genie or OpenAI’s efforts, aim to transcend data limitations and mimic human intuition.

While not necessarily the final answer, world models have already proven their value.

Elon Musk famously quipped: “You’re not driving by shooting lasers out of your eyes.” Six years later, Tesla’s FSD V13 demonstrated the power of first-principles thinking to Chinese EV startups.

Whether lidar is necessary remains debatable, but engineers who’ve tested FSD V13 often describe it as “leaving them speechless.”

That said, even if world models are theoretically superior, their initial versions may underdeliver.

When FSD Beta launched in October 2020, users reported significant regressions. Similarly, Musk’s transition from V12 to V13 (unifying highway and urban stacks) involved a rollercoaster of performance fluctuations.

4. Can You Share More?

Not really. @Harry_AD’s official video already speaks volumes.

Analyzing that clip alone, NIO is reclaiming its voice in the ADAS arena. Meanwhile, its emphasis on safety and accident rates aligns with its longstanding branding.

For perspective: While Chinese EV brands obsess over metrics, Musk often cites just one number—NHTSA’s “accidents per brand”—to promote Tesla’s Autopilot/FSD.

To wrap up:

NWM will define NIO in the autonomous driving era, and from its very first version, it’s already fundamentally distinct from past NOP iterations.