The final blockbuster headline in the autonomous-driving world for 2025 is here: L3 pilot programs have officially arrived.
Two Level 3-capable models are taking the lead in pilots across Chongqing and Beijing. One is the Deepal SL03, built on a pure-vision approach and running Changan’s self-developed NID intelligent driving system. The other is the BAIC Arcfox Alpha S6 (triple-LiDAR version), equipped with Huawei ADS.
What makes this round different from the earlier “local L3 permits” is a major shift in authority: these are the first L3 models approved at the national level by China’s Ministry of Industry and Information Technology (MIIT).
Almost overnight, L3 became the hottest topic again.
Huawei Intelligent Driving’s lead executive Jin Yuzhi posted a visibly excited line: “L3 is coming.” It’s understood that Huawei has already been running internal highway testing in Shenzhen using employee-purchase units of the Aito M9 and Maextro S800, logging over 1,000 kilometers.
That internal program differs from the public pilots in meaningful ways. It reportedly isn’t limited to a single lane, and it doesn’t confine testing to lower speeds. In many ways, it’s closer to what consumers imagine “real” L3 could feel like—than the two official pilot vehicles currently allowed on the road.
This wave of L3 excitement may sound intense, but it’s not the first “close moment” between L3 and automakers.
The first era arrived around 2017, when premium brands like Audi, BMW, and Mercedes unveiled L3 functions in Germany—sparking huge attention.
The second era came between 2018 and 2020, when multiple Chinese automakers—GAC, Chery, SAIC, Changan, and others—announced plans to mass-produce L3 vehicles. In practice, many so-called “L3 cars” amounted to little more than advanced self-parking.
Now comes the third era: 2025, arguably the year when the meaning of “L3” has changed the fastest.
Eight months ago, He Xiaopeng declared: “By the end of 2025, we will mass-produce L3-level smart driving capability.” Eight months later, the new Xpeng P7 appears with 2,250 TOPS of compute—but He Xiaopeng now says, “There’s only L2 and L4. Xpeng will not push L3 anymore.”
After eight years, L3 is no longer a single destination. It’s splitting into philosophies.
One camp argues you can skip L3 entirely, jumping from L2 toward L4.
The other insists L3 remains the most critical legal and regulatory step on the way to L5—something you simply can’t bypass when accountability is on the line.
Three routes to L3
The paths to L3 are more varied than most people think.
Under China’s national standard GB/T 40429-2021, L3 means the system continuously performs the full dynamic driving task within its operational design domain (ODD). In short: “conditional automated driving.”
The standard doesn’t spell out every restriction in detail, but for safety, current policy constraints place limits on speed, scenarios, and coverage for L3 pilots.
Route 1: Shrink the domain
If you can’t fully satisfy the rules, one approach is to shrink the ODD until it’s extremely manageable—a “conservative L3” strategy historically associated with German and Japanese automakers.
Starting in 2018, the major German luxury brands rolled out L3-capable models. Yet the “first global L3 model” narrative didn’t hold up cleanly: Audi’s A8, for example, never ended up delivering L3 at scale, and the team was ultimately disbanded.
Mercedes and BMW did mass-produce L3 systems, but only by compressing the operating scope to an extreme.
To make it work, they loaded vehicles with sensors—Mercedes’ S-Class Drive Pilot even includes a road-surface humidity sensor—yet the system still only operates in areas covered by high-definition maps.
Even then, it’s often limited to highways and capped at around 60 km/h. In real life, that version of L3 is difficult to call “useful.”
Route 2: Win by brute force—end-to-end redundancy
The second route is a full-stack “go big everywhere” approach—often associated with Huawei’s L3 ambition.
Huawei is expected to begin internal highway L3 testing on four models by year-end, including the Aito M9, Maextro S800, Arcfox Alpha S, and GAC Hyptec A800, all using Qiankun ADS 4.
This method is essentially an all-out push: more data, more redundancy, more safety layers.
Horizon Robotics founder Yu Kai has argued that one plausible way to reach L3 is to deploy L2++ at massive scale, accumulate enough statistical evidence, and only then “earn” the right to L3. Huawei’s path is close to that idea.
The core is a powerful data flywheel plus multi-layer safety redundancy. Huawei’s L3 internal-test configurations reportedly include at least four LiDAR units per vehicle. Meanwhile, Huawei has cited accumulated intelligent-driving data totaling 6.38 billion kilometers—a number few in the industry can match.
Earlier this year, Huawei’s push toward L3 already looked hard to contain. The “driver incapacitation assistance” feature in ADS 4—designed to guide the car safely to the roadside if the driver becomes unable to drive—already resembles “L3-like” capability in the eyes of many observers.
This route is extremely difficult to copy. For most players, it’s not a template—it’s a wall of resources.
Even without holding an L3 license directly, Huawei benefits from its ecosystem footprint. With ADS deployed across multiple Chinese automakers, the L3 momentum effectively lifts Huawei alongside the brands that hold the permits.
Route 3: Road-level foundation models
But a different group—especially new intelligent-driving players who emphasize AI foundation models—doesn’t want to wait.
That pressure is giving rise to a third route: road-level foundation models.
In Yuanrong Qixing founder Zhou Guang’s view, today’s L3 is built under existing technical frameworks and then constrained by rules—“can’t use in rain,” “can’t use at night,” and so on.
He argues L3 and L4 pull in different directions: one is rule-driven, while the other requires models that understand the world. The real bridge to both L3 and L4, he believes, is a foundation model that can learn road common sense at scale. If a foundation model trained on data at the “tens of millions” level becomes strong enough, L3 could become brief—or even skipped.
At the same time, L3 has one undeniable advantage: clearer responsibility boundaries, which can accelerate commercialization.
And that leads to a blunt reality: for many companies, L3’s business value currently exceeds its technical value.
L3 is often used more like an adjective than a capability.
You’ll hear “L3-level compute” to describe local effective compute above 2,000 TOPS. You’ll also see L3 used to sell cars—models like the Maextro S800 and Aito M9 are frequently marketed with lines like “the first L3 autonomous-driving model.”
The technology outcome is still uncertain—but the market reacts fast.
One day after the Arcfox Alpha S was included in the first batch of L3 pilots, BAIC’s stock jumped 10.01%.
Why L3 is moving slower than expected
Some industry insiders argue that, technically, L3 is a transitional stage. But in regulation, it’s not a “mere transition” at all.
In fact, L3 has consistently progressed slower than people assume.
Earlier this year, Xpeng, Huawei, Zeekr, and BYD all pointed to Q4 2025 as an L3 production window. Now, “mass-produced cars with L3-level capability” exist on paper—but true L3 rollout still sits at a distance.
At the same time, some players no longer treat “L3” as the main goal.
Xpeng has publicly signaled it will skip L3. Figures like Su Qing, Zhou Guang, and Yu Kai have also suggested that “L3 will be short.”
In a recent talk, Su Qing described Horizon Robotics’ new paradigm as breaking the split between L2 and L4—using a unified development framework, low-cost replication, and full-scenario generalization to deliver affordable “near-L4” systems.
The “disappearing L3” trend is driven by two forces: policy and user experience.
Policy: approval is slow, strict, and operationally heavy
L3 approval is a rigorous, lengthy process.
The two models currently piloting in Chongqing and Beijing were reportedly submitted more than a year ago—explaining why Arcfox’s pilot version is not running the newest ADS 4.
And in capability terms, both vehicles are restricted to single-lane operation. That means no lane changes while L3 is active—far from what consumers would consider usable L3.
In addition, safety-driven requirements make approvals particularly strict. Industry sources say obtaining L3 permits often requires automakers to work with an operating company as part of a joint application framework, with the operating entity taking legal responsibility during testing.
Even after qualification, each specific model still needs a test license. Today, the test fleet is largely to-B in nature, and broad to-C certification has not yet opened at scale.
This is also why the current pilot vehicles are widely viewed as non-mass-market versions.
Some consumers ask: didn’t Li Auto and Xpeng already announce “local L3 permits”?
To understand this, it helps to separate L3 progress into two tracks: local permits and the MIIT-led national pilot list.
From 2023 to 2025, more than 14 automakers reportedly partnered with local governments to obtain high-speed road testing permits for L3.
But those permits are limited to local jurisdictions. The MIIT-led list is the higher bar—and the one that matters most.
The formal program is the “Intelligent Connected Vehicle Access and Road Operation Pilot Consortium List.” Initiated in 2024 under MIIT leadership and approved jointly by four departments, it includes only nine consortiums.
Even so, the documents frame these pilots as a way to accumulate management experience and support future laws, regulations, and technical standards.
In other words: today’s pilots are designed to build institutional knowledge. Who will truly “break out” first with L3 remains undecided—and that uncertainty, plus strict long-cycle approvals, is fueling a new kind of intelligent-driving anxiety.
Experience: L3 doesn’t feel dramatically better for users
From the driver’s perspective, L3 doesn’t always deliver a clear leap.
In a restricted domain, L3 is less about technology and more about who takes responsibility. The accountability shifts from the user to the automated system.
But the driver still plays a “backup” role. Takeovers are still possible.
That means the driver can be hands-off, but not mind-off.
And this is the core contradiction: humans aren’t built for “no hands but full attention” for long periods. It’s a well-known limitation of human attention. That’s why newer standards require L3 systems to provide about 10 seconds for takeover and reaction time.
L3 is the most ambiguous—and the most caution-demanding—stage in the autonomy hierarchy. It is heavily policy-driven rather than purely technology-led.
And that reality clashes with a fiercely competitive intelligent-driving market—pushing some companies to explore skipping L3 entirely.
The “skip L3” logic, and the Tesla reference
The idea of skipping L3 inevitably circles back to Tesla.
“Tesla only has L2 and L4” has become a common justification among new players.
The SAE introduced the L1–L5 framework in 2014, but Elon Musk rarely emphasizes that hierarchy publicly. Tesla instead sells capability tiers under labels like AP, EAP, and FSD (supervised)—basic, enhanced, and full self-driving with supervision.
In Tesla’s own language system, high-level autonomy effectively collapses into two categories: FSD (supervised) and FSD (unsupervised).
The difference is simple: does the user need to supervise or not?
This “de-leveling” isn’t because Tesla doesn’t care. It’s because Tesla understands exactly what L3 implies.
One purpose is to avoid getting trapped in L3 approvals across countries. Many people assume overseas L3 regulation is looser. In reality, most jurisdictions treat automated driving with extreme caution.
Take the Czech Republic: starting January 1, 2026, the country is set to allow L3 unsupervised driving under rules that still require a 10-second takeover window, typically beginning with highways and expanding gradually.
Tesla’s FSD (supervised) is classified there as L2. To run L3, Tesla would need deep cooperation with authorities and would have to modify FSD to match local laws.
In practice, going L3 country-by-country is painfully slow—especially for a company with global ambitions. Even L2 global rollout isn’t trivial: FSD has only entered a limited number of countries so far, and L3 would be harder.
The second idea is more technical and more provocative: Tesla believes it has already shown that L2 can evolve into L4, at least architecturally.
Tesla’s Robotaxi uses an L4 system described as FSD (unsupervised), and it shares roots with the consumer-car FSD stack.
The hardware foundation is similar: both rely on HW4.0, use a pure-vision perception approach, deploy eight cameras, and run on Tesla’s self-developed chips.
On experience, He Xiaopeng has said after trying Tesla FSD and Tesla Robotaxi within a short window that the two experiences feel nearly identical—leading him to emphasize again that autonomy may jump straight to “near-L4 or full L4.”
Still, that’s the ultimate question. Robotaxi may prove the architecture’s ceiling is higher—but upgrading consumer L2 vehicles directly to L4 at scale remains a challenge even Tesla can’t fully guarantee.
Two philosophies—and why the split matters
Ultimately, today’s L3 disagreement reflects two competing routes to autonomy.
One is the Huawei-style approach: a gradual, full-stack path powered by data scale, redundant hardware, and tight end-to-end engineering.
The other is the Tesla-style approach: an AI-driven leap, betting that end-to-end systems and foundation models can compress the path from supervised driving to higher autonomy.
The emergence of this split is not a problem—it’s progress.
When the industry debates L3 for the third time at scale, the center of gravity has shifted. The conversation is no longer dominated by HD maps and traditional L3. It’s increasingly about end-to-end systems plus large models as the route to higher autonomy.
Whichever route becomes mainstream, the debate itself is what pushes the industry closer to what it truly wants: L4.
