The assisted driving industry is at a sensitive turning point.
On one hand, the push to bring the technology to more drivers is finally moving from talk to mass production, reaching mainstream vehicle segments. On the other hand, more computing power, larger models, and louder technical claims are advancing in parallel, with goals aimed at Level 3 and even Level 4 automation. It is a moment of opportunity, and a moment of pressure.
After several rounds of industry shakeouts, QCraft has emerged as one of the survivors. It did not start with the strongest position, and it did not win by raising the most money. Instead, it has repeatedly made measured calls at key moments, and most of them proved right.
When the industry was still fixated on Level 4 autonomous driving, QCraft was among the first to step back and shift toward the mass production of Level 2 assisted driving software. It also moved early to cut the technical and organizational baggage that often came with close ties to autonomous driving startups.
Later, as much of the automotive sector standardized on Nvidia chips and the computing arms race accelerated, QCraft took a different path. It partnered with Horizon Robotics, built its software around the Journey chip series, and moved earlier than many peers into a part of the market where cost and efficiency matter more.
As a software and algorithm provider, QCraft has kept close but independent relationships with both automakers and chipmakers. Yu Qian, QCraft’s co-founder, chairman, and CEO, told 36Kr that its early, deeply integrated work with Li Auto produced two things it still relies on: a repeatable way to deliver to mass production, and a disciplined approach to defining products.
Those capabilities are scarce in early-stage assisted driving startups, and they are easy to miss when the conversation is dominated by big technical promises. Many companies that later exited were not beaten on algorithms. They were pushed out because they could not deliver reliably and at scale over time.
But close alignment with an automaker has risks. As automakers build stronger in-house teams, suppliers can quickly lose influence or get pushed to the margins. Nearly every assisted driving startup eventually has to deal with that tradeoff.
QCraft does not see that as a crisis. Yu framed it as a healthy sign. “In-house development is not a bad thing,” he said. “Automakers should have their own R&D capabilities. They need to develop the judgment to know what is good and what is not.” In some cases, he added, QCraft is even willing to help automakers build those in-house capabilities.
That same idea, staying close without becoming dependent, also shapes how QCraft works with Horizon.
Although QCraft joined Horizon’s ecosystem early, it avoided deep binding. It has continued adapting its solutions for Nvidia and Qualcomm platforms, keeping them portable across different chip architectures.
“We’re an important ecosystem partner, but we remain independent,” Yu said. “We don’t rely on Horizon’s algorithms. We only use its chips and general toolchains. All of our underlying algorithms, simulators, and libraries are developed in-house, so our roles don’t conflict.
A string of choices like these has helped QCraft stay in the game. The company says cumulative installations of its assisted driving systems in passenger vehicles have surpassed one million units. By 2026, it expects the number of mass produced vehicle models it supports to exceed 50, with almost all equipped with urban navigate-on-autopilot (NOA) features.
Still, the shakeout is not over. The field remains crowded. Horizon has entered the market with its HSD solution. Automakers, even when they buy from external providers, continue to pursue in-house development. Yu does not expect the market to settle around a single dominant player. He thinks it will look more like engines or batteries, with four to five leading companies.
QCraft is trying to be one of them.
Looking ahead, it is also pushing into new areas. On one front, it is expanding its Level 2 lineup, offering three tiers of solutions across different chip platforms. On the other hand, it is increasing investment in Level 4 technology, including new use cases such as autonomous logistics.
Its next step may be a move toward the capital markets.
The following transcript has been edited and consolidated for brevity and clarity.
36Kr: You recently tried Tesla’s FSD (Full Self-Driving) solution. How did it feel?
Yu Qian (YQ): Aside from occasionally misjudging parking garage exits or missing a turn, whether inside parking structures or on surface roads, it can basically drive itself. The experience is very smooth. It can even chat with you, which makes the interaction feel quite natural.
FSD version 14 is genuinely very good. Our current performance is probably closer to version 12. I don’t think Chinese automakers or technology companies lack capability. With a lag of six months to a year, they can definitely catch up to that level.
36Kr: What is the main advantage of QCraft’s end-to-end solution?
YQ: Our core advantage is achieving the best possible experience with limited resources. With a single Journey 6M setup, our experience is better than many dual Orin X solutions already in mass production. I won’t name specific brands.
End-to-end approaches use computing power more efficiently. You don’t need to stack multiple models. One model can cover many features, making it more efficient than two-stage approaches. We don’t blindly pile on computing power or parameters. We focus on the core needs of 90% of users and avoid flashy, show-off features.
36Kr: What are QCraft’s plans for vehicle adaptation and solution deployment in 2026?
YQ: We expect to add more than 50 new vehicle models in 2026, with over 30 already in production. Domestically, the main solution is Horizon’s Journey 6M, accounting for more than half of deployments, alongside Journey 6E and Qualcomm-based solutions. Overseas, Nvidia and Qualcomm are the primary platforms.
We’ve divided our product matrix into three tiers.
- The Air tier covers highway NOA plus active safety, suitable for internal combustion and electric vehicles priced below RMB 100,000 (USD 14,000), using air-cooled systems.
- The Pro tier focuses on urban assisted driving for the RMB 100,000 segment, with around 200 TOPS (tera operations per second) of computing power, a cost of several thousand RMB, 11-volt sensors, and optional LiDAR (light detection and ranging).
- The Max tier offers more than 500 TOPS of computing power and targets extreme urban NOA. The chip model has not been disclosed, with mass production planned.
36Kr: How is QCraft laying out its work across end-to-end systems, world models, and vision-language-action (VLA) models?
YQ: End-to-end is our main focus, while we also advance world models and VLA models. End-to-end solutions are still in the pre-research phase and are expected to reach mass production in 2026.
VLA is essentially a form of end-to-end. The core idea is enabling the system to both know what it is doing and understand why, improving generalization. It is not about translating images into language and then outputting actions. We plan to deploy it in combination with robotaxi and Level 4 businesses.
World models are currently used mainly for cloud-based virtual training to solve scenario reproduction challenges. Their priority is much higher in the cloud than on the vehicle side, where applications are not yet mature.
36Kr: Why are world models not yet mature on the vehicle side?
YQ: Today, world models cannot replace road testing at scale. They may account for only 10–30% of training value. If one day they could fully replace road testing, the development curve of autonomous driving would shift from gradual to sharply upward.
A true world model should capture all the physical laws of the real world within a virtual environment. That requires enormous computing power, which is why world models matter far more in the cloud than in the vehicle.
36Kr: How do you approach vision-language-action models? Do you translate visual information into language and then output actions?
YQ: That’s a misunderstanding. We don’t do it that way. Translating images into language loses a lot of detail. Driving often relies on intuition, and there’s no need for that extra step.
The role of language is to extract and abstract driving behavior, like a coach injecting knowledge so the model can generalize and learn by analogy, closer to how humans learn, rather than simply translating signals.
36Kr: What kind of experience do you want vision-language-action to deliver?
YQ: Right now, our priority is enabling the majority of consumers to actually use assisted driving in cities with a good experience. That means making tradeoffs on flashy demonstrations.
If 90% of people can get 90% of the experience, VLA can help achieve that. At very high computing power levels, it can deliver incremental gains, but the marginal value may not be that large. Our current pre-research positions it more for robotaxi and Level 4 use cases, where deeper generalization and scenario understanding matter, and that will take more time.
36Kr: Do these technologies mean QCraft is hitting a computing power ceiling?
YQ: That’s a question of resources. If resources were truly the main factor, DeepSeek wouldn’t exist.
We admire DeepSeek. With limited computing power and very small resources, it still delivered strong results through innovation. That’s how real value is created for customers.
36Kr: Li Auto recently rolled out its AD Pro 4.0 update. How did QCraft initially start working with Li Auto?
YQ: We originally provided planning and control solutions to another automaker in Guangzhou. But even then, we already had full-stack capabilities based on the Journey 5 platform. We showed Li Auto a demo, and nearly 100 people from its team tested it. In the end, it chose our solution.
We also learned a great deal from Li Auto about mass delivery, how to define products properly, how to ensure delivery timelines, and how to manage the entire product definition process. During that period, our team essentially formed a complete operating playbook, which was extremely important.
36Kr: At its peak, how large was the team working on the Li Auto project?
YQ: Around 100–200 people.
36Kr: Is there always tension between suppliers and automakers pursuing in-house development?
YQ: We’re not worried about automakers doing in-house development. Many of the automakers we work with are doing it, including Geely, Li Auto, and GAC. We never claim that assisted driving will all end up the same, or that it is not worth doing in-house.
Our attitude is to be good and to want them to be good. From our perspective, we’re willing to help automakers build their in-house capabilities. Whether they ultimately succeed depends on the product and the experience. Only good experiences win through real competition.
In-house development is not a bad thing. Automakers should have their own R&D capabilities. They need to develop the judgment to know what is good and what is not
36Kr: Has end-to-end technology raised the bar for in-house development?
YQ: Absolutely. The difficulty is much higher. On a global scale, most automakers still haven’t achieved it. It takes a long time, and the challenges span talent, computing power, data, and more. Success in this area is inherently a low-probability event.
36Kr: After end-to-end tech gained traction, many companies that were previously quiet suddenly reappeared. Why?
YQ: They weren’t starting from zero. Many have nearly a decade of accumulated experience, often coming from Level 4 backgrounds. They simply found the right optimization direction and a better match between technology and vehicle deployment.
36Kr: How does QCraft view its relationship with Horizon Robotics?
YQ: I don’t think there’s a contradiction. We don’t actually buy chips. Horizon’s chips are supplied directly to automakers and tier one suppliers. What we mainly provide is software, which makes us a software supplier.
Our relationship with Horizon is more cooperative than competitive. We’re an important ecosystem partner, but we remain independent. We don’t rely on Horizon’s algorithms. We only use its chips and general toolchains. All of our underlying algorithms, simulators, and libraries are developed in-house, so our roles don’t conflict.
At the same time, we’re also ecosystem partners of Nvidia and Qualcomm. Domestically, we promote Horizon-based solutions. Overseas, we focus on Nvidia and Qualcomm, adapting based on market demand.
36Kr: When delivering projects, does QCraft usually act as a tier one or tier two supplier?
YQ: It’s flexible. Sometimes we’re a tier one supplier, sometimes a tier two supplier.
That distinction doesn’t determine who has more influence. What matters is who creates the scarcest value. That said, in the automotive sector, automakers ultimately hold the strongest bargaining power.
36Kr: Is there a real difference between being an ecosystem partner and a regular partner?
YQ: Fundamentally, not much. It’s mostly a difference in external positioning. Every ecosystem is extremely competitive. We may get earlier access to chip versions or better localized support, but ultimately the core still comes down to our own technical optimization.
36Kr: How do you assess assisted driving penetration and commercialization?
YQ: Globally, assisted driving penetration is currently below 5%. Over the next few years, it could rise to 50%, entering a tenfold growth cycle. This is closely tied to electrification. Electrification is the foundation of intelligence. Without large-scale electrification, assisted driving cannot truly take off.
A positive value cycle will emerge once penetration reaches 30–40%. At that point, advantages in accident and minor collision rates will become clear, insurance premiums will begin to reflect that value, automakers will be more willing to pay for high-quality solutions, and weaker players will be eliminated. I think this inflection point could arrive within one to two years.
36Kr: How do you view assisted driving moving into lower-priced vehicle segments?
YQ: Vehicles in the RMB 100,000 range should eventually be equipped with urban NOA. A few years from now, they may even approach Level 4 assisted driving capability. Technology diffusion is inevitable. Ordinary consumers also deserve safe, comfortable, and convenient experiences.
Ultimately, assisted driving will become standard across vehicles.
36Kr: What kind of competitive landscape do you expect China’s assisted driving market to settle into?
YQ: It won’t end up with only one or two companies. Automakers support multiple suppliers to maintain supply chain control and healthy competition. Assisted driving will follow the same logic, and the market will likely retain four to five leading players.
The upstream and downstream parts of the industry chain are inherently competitive. Automakers hold the strongest bargaining power and won’t allow a single supplier to monopolize the market.
36Kr: QCraft often talks about focus and tradeoffs as the core of its strategy. How did that mindset take shape?
YQ: Strategy is about focus and tradeoffs. Among the things you want to do, the things you can do, and the things you should do, you choose what you should do.
We focus on software algorithms and avoid hardware manufacturing, which is not our strength. We integrate supply chain resources through partnerships, and that is our core strategy.
36Kr: So QCraft has no plans to move into chips or other hardware?
YQ: For now, we won’t enter large-scale manufacturing of sensors or chips. Those fields already have mature players. The chip industry requires massive investment, long development cycles, and strong economies of scale. Automotive-only chips are particularly hard to make profitable.
36Kr: What about embodied intelligence and robotics?
YQ: In the long term, yes, we will move in that direction. Assisted driving itself is a major application of physical artificial intelligence. It is essentially embodied intelligence on four wheels. But for now, our priority is assisted driving and autonomous logistics. We want to win the current battles first. Once those businesses are stable, we’ll gradually expand into robotics.
36Kr: How do you see the timeline for Level 4 assisted driving and robotaxis?
YQ: Level 4 technology and robotaxis will take a long time to materialize, especially in China, where the environment is complex, labor costs are low, and regulations are restrictive. But we’ve never abandoned that long-term goal.
Our plan is to first break through in autonomous logistics. We’re already operating in multiple cities in partnership with Chery. After January, we’ll release more information and then gradually push toward robotaxi deployment.
36Kr: What is the current progress and roadmap for the autonomous logistics business?
YQ: We’re working with Chery to develop autonomous logistics vehicles built to automotive-grade standards, balancing stability, reliability, and cost. These vehicles are already operating in multiple cities.
Our goal is to achieve large-scale mass production in logistics scenarios first, then move toward robotaxis. We expect to reach a meaningful scale within two to three years.
This is essentially a form of down-the-stack competition. We’re fully leveraging mass production capabilities. Our supply chain and quality control are stronger than those of many existing players. Chery handles manufacturing, which gives us more stable automotive-grade quality and lower costs.
36Kr: Some players have already locked up logistics hubs and sites. Do you feel late to the market?
YQ: Not at all. This industry is only just getting started. Leading players today have deployments of only tens of thousands of vehicles. China has around 20 million logistics vehicles in total. The future market will be in the hundreds of thousands, or even millions.
At that scale, stability and reliability become paramount. Our advantage lies in mass production and automotive-grade quality. That is our core competitiveness.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Xiao Man for 36Kr.