Skip to content
China AI DispatchChina AI news, every weekday

Companies / BYD

Automaker and AI chipmaker

BYD

State of play as of

The world’s largest EV maker, now running its own self-driving world model on a chip it designed itself.

BYD (比亚迪) is the world’s largest electric-vehicle maker, and through 2026 it revealed a second identity almost no one writing in English had noticed: an in-house AI operation built to own the full self-driving stack, the driving data, the world model and the silicon. Its first published paper, HyWorldVLA, posted what BYD reports as the top score on the public NAVSIM v1 self-driving benchmark, and its in-house Xuanji A3, which BYD calls China’s first 4nm automotive-grade driving chip, is meant to run that model in its own cars. The open questions are the ones any leaderboard raises: a benchmark record is not a shipping product, BYD has named no deployment date, and a public benchmark can be over-fit. We track what BYD has built, what it claims, and where the caveats sit, with dates and sources.

Fast facts

What it is
BYD (比亚迪) is the world’s largest electric-vehicle maker, now developing its own self-driving models and automotive AI chips in-house.
AI debut
HyWorldVLA, BYD’s first published AI paper, is a driving world model that scored 90.59 on the public NAVSIM v1 self-driving benchmark, which the paper reports as the highest score posted on it. The figure is from BYD’s own paper.BYD paper, via ithome and Pandaily, via CAD
Its own chip
Xuanji A3, which BYD calls China’s first 4nm automotive-grade driving chip, rated about 700 TOPS per chip and up to 2,100 TOPS with three linked, at the highest automotive safety level.BYD, via CAD
The stack
BYD owns the driving data, the world model and the chip, and says it is the first company doing all three at the scale of millions of vehicles a year.CAD analysis
Manufacturing scale
BYD has produced more than 10 million EVs and builds on the order of 30,000 vehicles a day, the supply-chain base it says underpins the AI push.BYD EVP Li Ke, via TechNode, via CAD

The state of play

  1. BYD’s AI research team published its first paper, describing HyWorldVLA, a driving world model that scored 90.59 on NAVSIM v1, the public benchmark used to grade self-driving policies, which BYD’s paper reports as the highest score posted on it. The design pairs a pixel-level branch that predicts the road ahead frame by frame with a faster branch running in a compressed abstract space; BYD trains the two together and runs only the fast branch when the car drives. Removing either branch lowers the score, to 87.50 without the pixel supervision and 89.91 without the abstract layer. The result BYD stresses is bad weather: across 655 rain-and-fog scenes the model scored 86.87, more than 19 points above the strongest baseline. Two caveats travel with it. The model has not driven a paying customer on a real street, BYD has named no ship date, and a public benchmark can be over-fit.

    BYD paper, via ithome and Pandaily, via CADRead the issue →

  2. LatePost reported BYD plans to put its own Xuanji A3 chip in a production Denza in 2027, a 4nm part rated above 700 TOPS per chip and above 2,100 with three linked, supporting L3 and L4 driving. Chairman Wang Chuanfu’s framing was the quotable part: the first half of the EV race was about the battery, he said, the second half is about the chip.

    LatePost, via CADRead the issue →

  3. The Pentagon’s annual Section 1260H update added BYD to its list of Chinese military companies, alongside Alibaba, Baidu and Unitree, part of a broader sweep of Chinese EV and battery makers including Nio and CALB, with a 30 June cutoff. The listing is a reputational and procurement signal rather than a sanction; it does not by itself bar US sales.

    US DoD Section 1260H, via CADRead the issue →

  4. BYD confirmed it is developing humanoid robots. EVP Li Ke said automotive AI and robotics share common foundations, and that BYD would build its own robots while running an open platform for partners, with its roughly 3,700-store dealer network as a possible sales channel. The argument BYD makes is scale: it has produced 10 million EVs and builds about 30,000 vehicles a day, the same supply-chain discipline that actuators, motor drivers and joint assemblies need at industrial volume.

    BYD EVP Li Ke, via TechNode, via CADRead the issue →

  5. BYD detailed Xuanji A3, a 4nm car-grade AI processor it built entirely in-house and puts on the same process tier as Nvidia’s Thor, delivering 2,100 TOPS across three dies with, by its own account, 20 percent lower power per unit of compute and higher utilization than a general-purpose GPU adapted for cars. Founder Wang Chuanfu offered uncapped liability for accidents under city navigation, a promise he tied to BYD controlling the battery, powertrain, sensors and now the compute.

    BYD launch event, via QbitAI, via CADRead the issue →

Our reporting on BYD

  1. The Carmaker

    BYD, the world’s largest EV maker, published its first AI paper. HyWorldVLA, a two-branch driving world model, scored 90.59 on the public NAVSIM v1 self-driving benchmark, which BYD’s paper reports as the top score posted on it, and held a 19-point lead over the strongest baseline across 655 rain-and-fog scenes. Paired with the in-house Xuanji A3, which BYD calls China’s first 4nm automotive-grade driving chip at about 700 TOPS, it gives BYD its own data, model and silicon stack. The benchmark record is not yet a shipping product and BYD has named no deployment date.

  2. The Harder Half

    Cambricon became China’s first pure-play AI chip designer worth over a trillion yuan. The same report shows the distance left: a 2.9% share of the cards China bought in 2025, against Nvidia’s 55%.

  3. The List

    The Pentagon’s annual Section 1260H update added Alibaba, Baidu, BYD and Unitree to its list of Chinese military companies, with a 30 June cutoff. BYD was folded into a broader sweep of Chinese EV and battery makers including Nio and CALB.

  4. The Sprint

    BYD confirmed it is developing humanoid robots, arguing automotive AI and robotics share the same foundations and that its manufacturing scale, on the order of 30,000 vehicles a day and a roughly 3,700-store dealer network, is an edge no pure-play robotics startup can match. The same week Xiaomi’s robot swept two research championships and Xpeng set a Q4 mass-production deadline.

  5. The Commissioning

    Three Chinese teams released competing embodied world models in the same 72 hours, and BYD detailed Xuanji A3, a 4nm car-grade AI chip built in-house that it puts on the same process tier as Nvidia’s Thor, rated at 2,100 TOPS across three dies. Founder Wang Chuanfu offered uncapped liability for autonomous-driving accidents, a promise he tied to controlling the full stack.

These link to the full issues on the newsletter. New pieces go out in the daily first.

Common questions

What is BYD’s AI strategy?

BYD (比亚迪), the world’s largest EV maker, is building the full self-driving stack in-house: the driving data from its own fleet, a world model of its own, and its own automotive AI chip. In 2026 it published its first AI paper, HyWorldVLA, and shipped the Xuanji A3 driving chip, so it now owns the data, the model and the silicon rather than buying any layer from a vendor. The bet is vertical integration, and it means BYD has to be world-class at chips and at models and at cars all at once.

What is HyWorldVLA?

HyWorldVLA is BYD’s first published AI model, a self-driving world model that scored 90.59 on NAVSIM v1, a public benchmark for driving policies, which BYD’s paper reports as the highest score posted on it. It uses two branches, a slow pixel-level one that predicts the road frame by frame and a fast one running in a compressed abstract space, trained together and with only the fast branch used when the car drives. BYD highlights a bad-weather result of 86.87 across 655 rain-and-fog scenes, more than 19 points above the strongest baseline. The score is from BYD’s own paper, and a public benchmark can be over-fit.

What is the BYD Xuanji A3 chip?

Xuanji A3 is BYD’s in-house automotive driving chip, which BYD calls China’s first 4nm automotive-grade driving processor and puts on the same process tier as Nvidia’s Thor. BYD rates it at about 700 TOPS per chip and up to 2,100 TOPS with three linked, at the highest automotive safety level, and reports that as a specialized NPU it runs at lower power and higher utilization than a general-purpose GPU adapted for cars. LatePost reported BYD plans to put it in a production Denza in 2027 for L3 and L4 driving. The performance and efficiency figures are BYD’s own.

Does BYD’s self-driving beat Waymo or Tesla?

BYD’s HyWorldVLA posted the top reported score on NAVSIM v1, a public benchmark, which means it planned safer and more complete driving than previously published systems in that benchmark’s scenarios. That is a leaderboard result, not a road-deployment claim: the model has not carried a paying passenger, BYD has named no ship date, and a public benchmark can be over-fit. Where BYD does stand apart from Waymo and Tesla is integration. Waymo buys its compute from Nvidia and Tesla designs its own chip but sells far fewer cars, while BYD is the first to own the data, the model and the silicon at the scale of millions of vehicles a year. On real-world driving, no independent comparison exists yet.

Is BYD building humanoid robots too?

Yes. In June 2026 BYD confirmed it is developing humanoid robots, with EVP Li Ke arguing that automotive AI and robotics share common foundations and that BYD’s manufacturing scale, roughly 30,000 vehicles a day and a 3,700-store dealer network, is an edge over pure-play robotics startups. BYD said it would build its own robots while also running an open platform for partners. As with its cars, the thesis is that owning the supply chain, from actuators to compute, is the advantage.

Every dated fact above traces to a named source, and where it comes from our reporting, to the dated issue that carried it. Where a claim is self-reported or disputed, we say so.

Subscribe

The stories English coverage misses, every morning

I scan 100+ Chinese-language sources every day and write up the China AI stories English coverage misses. It goes out every weekday. Free.

Free. One email each weekday morning. Unsubscribe anytime.