Driver assistance: hardware and software

Driver assistance in XPENG’s European models runs on one of three hardware platforms. Bosch supplies a camera module, NVIDIA an Orin-X chip, and XPENG developed the Turing chip itself. Which one your car has depends on model, trim and model year, and it decides which features you can get.

This page is the reference half of the subject: what the hardware is, what the software is called, and where the European cars part company with the Chinese ones. How the systems behave on the road, and the habits that make them useful, sit in the practice guide, which follows later. Written August 2026, with XOS 5.9.5 as the current international software.

Images: XPENG

Jump to: Three platforms · Bosch · NVIDIA Orin-X · XPENG Turing · Platform by model · The software on top · Second-generation VLA · The research behind it · China and Europe are converging · Q&A · Sources

The three platforms

XPENG buys the assistance hardware from Bosch on the cheaper cars, from NVIDIA for the G9, the P7 and most G6s, and for the newest models makes it in-house. The split runs straight through the middle of two model ranges. Two G6s in the same Danish showroom can be on different platforms (Bosch or Orin-X), and so can two L03s (Bosch or Turing), which is why an equipment list is worth more than a badge when you are working out what a car can do. The three are not the whole story: the Norwegian G3 of 2020 and the original P7, on sale in Norway from October 2021, both predate them and run on different hardware.

Bosch: the front camera module

It sits in the G6 MY25 Standard Range / Air, and on the L03 in every trim except the Ultra: Standard Range, Long Range and AWD Performance, the PowerX Long Range included. The European price lists describe the hardware in one line, a front-integrated Bosch camera system for ACC and LCC, and leave it at that.

The German L03 equipment list runs to twenty named functions on Bosch hardware: adaptive cruise (ACC), lane centering (LCC), speed assist, sign recognition, automatic parking, remote park, remote forward/back, emergency braking, fatigue monitoring, forward collision warning, front-cross-traffic warning, blind-spot detection, door-open warning, the four lane-safety functions, rear collision warning, rear-cross-traffic warning and intelligent high beam. Everything the EU requires of a new car is there, and most of what an owner uses daily.

The limit shows at the top of the stack. XPENG’s own 5.9.5 release note says custom APA parking is not available on the G6 MY25 Standard Range / Air, and gives the reason directly: that variant has neither an NVIDIA Orin-X nor an XPENG Turing chip. XPEDAL with full stop is another function that varies by car. Beyond those two, the difference is undocumented. XPENG has never published a Bosch part number, a compute figure, or a function-by-function comparison, so any complete list of what a Bosch car misses would be guesswork.

XPENG counts Bosch among its long-term European suppliers. The company stood on the partner wall at Brand Day in Munich on 16 July 2026, one of twenty-three names alongside NXP, Infineon, ZF, Valeo and Michelin. Why this particular part is bought in when XPENG makes its own chips, the company has never explained. Two reasons suggest themselves: XPENG’s own Turing chips are a scarce resource that goes to the cars meant to run NGP, and an entry trim has a price to hit. The pattern fits either: Bosch sits under the cheapest trims, XPENG’s own or NVIDIA’s silicon above them.

NVIDIA Orin-X: the platform from 2022

One Orin-X drives assistance in the G9 through every model year Europe has had, MY23 to MY25, in the P7 from the 2023 facelift, in every trim of the original G6 (MY24), and in the G6 MY25 Long Range and Performance. The newest car on the platform is a 2025 model; everything from 2026 runs Bosch or Turing - the L03, the P7+ and the X9. That makes Orin-X the outgoing one of the three - still in cars you can buy new today, but not in the ones XPENG is drawing now. Bosch and Turing are both live choices, one for the cheaper trims and one for the top.

The NVIDIA relationship is older than the export push. The China P7 of April 2020 was the world’s first production car on NVIDIA DRIVE AGX Xavier - 30 TOPS at 30 watts - and the in-house ADAS story XPENG now tells through Turing starts on that chip. When the G9 launched at home in September 2022, its top versions paired two Orin-X chips - 508 TOPS - with dual LiDAR and 31 sensors for XNGP, while the export car that later reached Denmark kept one chip and no LiDAR.

The sensor set around it is the largest XPENG has fitted in Europe: five millimetre-wave radars, twelve ultrasonic sensors, four surround cameras, seven high-perception cameras and a cabin camera. XPENG Denmark’s xDIVE explainer of December 2024 adds the detail the price lists skip: the forward view is an 8-megapixel binocular camera pair, the high-perception cameras run at up to 3 megapixels, and the cabin camera reads the driver in infrared. The P7 Wing Edition carries fourteen ultrasonic sensors instead of twelve.

No European configurator prints a TOPS figure for these cars, but NVIDIA publishes one for the chip: 254 TOPS, which its developer material states as INT8 across the GPU’s tensor cores and the deep-learning accelerators. Orin-X is NVIDIA’s own name for the part, not an importer’s shorthand. The same documentation puts the platform on ISO 26262, the functional-safety standard XPENG’s cars are homologated against, with a dedicated safety area on the chip that NVIDIA calls a safety island.

The obvious sum does not work, though. XPENG quotes its Turing chips as effective compute and has never named a datatype for the 750 TOPS, so one Orin-X against one Turing sets a measured number beside an unspecified one.

The hardware has aged well. A G9 delivered in 2023 shipped on XOS 4.x, was documented under a manual describing that software, and drives on 5.9.5 today. From 4.x to 5.x on one chip, with the assistance functions growing across all of it.

XPENG Turing: the chip they developed themselves

The Turing chip reached European cars with the P7+ and the X9 in 2026; the price lists write it as Turing XP5.

Turing is the chip XPENG is betting on. The tape-out, when the finished design goes to the foundry, came on 23 August 2024, and at AI Day that November the company described a 40-core processor on a domain-specific architecture for neural networks, running models of up to 30 billion parameters and delivering, in its own words, the power of three high-performance chips in one. Its own chip pages fill in the anatomy: two in-house NPUs, two independent image processors for darkness and backlight, and an integrated safety island certified to ISO 26262 ASIL-D and AEC-Q100 Grade 2. The 750 TOPS in the configurators is the retail version of the compute claim; no independent benchmark of the chip has been published.

Three European models have it, and the G9L is on the way. The P7+ and the X9 carry one each. The L03 Ultra carries three chips. Two of them handle the driving at 750 TOPS each, so 1,500 TOPS, and the third handles the cabin. The name is not arbitrary: in China the chip count is a trim step, where Max carries one, Ultra-SE two and Ultra three, and the L03 is the first European car to wear Ultra. GX and G9L are sold that way for 750, 1,500 and 2,250 TOPS of effective compute, and the Robotaxi version of the GX takes up to four for 3,000 TOPS with dual-redundancy hardware. Wherever a Chinese car carries three, the split between driving and cabin is the same, and the 2,250 TOPS on the posters counts all three.

The sensor set shrank as the compute grew. A Turing car runs three millimetre-wave radars instead of five, twelve ultrasonic sensors, and eleven cameras on the P7+ or twelve on the X9. And, as on every export XPENG, no LiDAR.

The same silicon runs in three product lines. At Brand Day in Munich XPENG presented AI cars, flying cars and humanoid robots. The flying vehicles belong to its subsidiary Aridge. The IRON humanoid robot runs on three Turing chips delivering 2,250 TOPS, allowing it to run the physical AI model locally on-device without teleoperation; it is scheduled for global release in 2027. The chip has its first outside customer too: at AI Day in November 2025 XPENG named Volkswagen the launch customer for VLA 2.0, with the Turing chip nominated for the group’s cars.

Platform by model

European trims, from the Danish and German price lists captured in July 2026.

Car	Hardware	Sensors	Software layer
G9 MY23-MY25	1 x NVIDIA Orin-X	5 radar · 12 ultrasonic · 4 surround + 7 high-perception cameras · 1 cabin	XPILOT ASSIST 2.5
P7 facelift 2023 · MY24	1 x NVIDIA Orin-X	Same set; Wing Edition has 14 ultrasonic	XPILOT ASSIST 2.5
G6 MY25 Long Range · Performance	1 x NVIDIA Orin-X	Not itemised	XPILOT ASSIST 2.5
G6 MY25 Standard Range / Air	Bosch front camera module	Not itemised	XPILOT ASSIST 2.5
L03 MY26 SR · LR · AWD Performance	Bosch front camera module	Not itemised	XPILOT ASSIST 2.5
L03 MY26 Ultra	3 x Turing; 2 for driving = 1,500 TOPS	Not itemised	XPILOT ASSIST 2.5 · NGP / VLA 2.0 planned Q1 2027
P7+ MY26	1 x Turing XP5	3 radar · 12 ultrasonic · 11 cameras	XPILOT ASSIST 2.5
X9 MY26	1 x Turing XP5	3 radar · 12 ultrasonic · 12 cameras	XPILOT ASSIST 2.5

The software on top

XPILOT ASSIST is a marketing name, not one system. The Bosch trims run the supplier’s own system, while the Orin-X and Turing cars run XPENG’s. The name is the same while everything under it varies. The G6, G9, P7, P7+, X9 and L03 carry exactly the same label in the Danish, Dutch, Norwegian and Swedish stores: three different hardware platforms, model years 2023 to 2026, one number. It marks neither a software release nor a hardware generation. The software under it has moved a whole generation on, and 2.5 has not budged. The number is a leftover. 2.5, 3.0 and 3.5 belonged to one car each in the G3i, P7 and P5 generation, and 3.5 reached the P5 in 2021. XPENG’s model pages today say XPILOT ASSIST with no number at all. Only the price lists still carry 2.5, and there it sits on everything from a Bosch camera module to three Turing chips.

Above XPILOT ASSIST sits NGP, and so far it has reached exactly one European equipment list: the L03 Ultra’s, where XPENG has NGP on VLA 2.0 planned for the first quarter of 2027. The wider roadmap, approval work included, is set out on the G9L page.

A version number can mean two things. XOS is the cockpit and vehicle software, with its own version numbers from 1.0 in 2018 to the 6.x generation now shipping in China; the XOS version history lists them all. The driving stack rides along inside those releases without sharing their numbering, which is why an OTA can change assistance behaviour while no assistance version changes name. The oldest name in the family is XPILOT 3.0, tied to XOS 3.0 in China in 2021 and never sold outside it.

Second-generation VLA

VLA stands for vision-language-action: one model that takes camera input and instructions and puts out driving decisions, in place of a chain of separate perception and planning modules. XPENG calls it VLA 2.0 version 6.3.0, where 6.3.0 is the XOS release it arrives with, and describes it in English as VLA 2.0’s first major upgrade, while the Chinese store lists it as second-generation VLA. The launch posters shorten the number to the 630 version, and the G9L is the first car in the world to get it. The Chinese rollout starts in September 2026 with the full model on Ultra and Ultra SE.

Four pieces carry names of their own:

• Infini-VLA - up to thirty seconds of history feed the driving decision. XPENG frames it as adding time to a model that previously had only space
• X-Foresight - a world model running on the car that sketches where other road users are likely to go about six seconds ahead. Flow Matching is the technique named on the same slide
• Streaming inference - perception, reasoning and action run together rather than in sequence. This is the change behind the 300% faster end-to-end response XPENG quotes
• Master Agent - VLA joined to a vision-language model on XPENG’s own Omni multimodal base. It takes a natural-language instruction, automatically breaks it into subtasks, and passes them to separate agents for driving, chassis, cabin and body control. In practice, this enables actions such as asking the car by voice to pull over and park or find the nearest available bay while driving on the assistant. A Mixture of Tasks (MoT) architecture dynamically routes tasks like urban driving, campuses and parking to dedicated expert models to prevent interference

The scale claims are XPENG’s own measurement on the Chinese car: 3.5 times the on-device parameters of the shipping version, which the company scales at more than fifteen times a mainstream VLA model, navigation mistakes down 50%, and a twentyfold lift in what it calls multi-dimensional safety. The cloud model behind it has a published size - 72 billion parameters, named at AI Day in November 2025 - while the on-device count has not. And the “Robotaxi-like L4-level experience” on the same slides is primarily marketing wording - no L4 approval exists for these cars anywhere yet.

For a European buyer the hardware floor matters more than any of those figures. Outside China, XPENG’s own product split points to at least two Turing chips to get NGP. The P7+ and X9 have one Turing chip each, the G6 and G9 have none, and the L03 Ultra is the only European model with two Turing chips dedicated to driving and VLA 2.0 on the equipment list. For single-chip cars the answer already exists in China: second-generation VLA Lite, a distilled edition. It rolls out from September 2026 with XOS 6.3.0 to Max trims with one Turing chip, and the G9L Max gets it first. If the P7+ or X9 ever get NGP, that distilled version is what they will likely receive.

Lite is the full model with a smaller vision encoder. He Xiaopeng explained the choice in the clip XPENG posted on Weibo on 29 August: halving the parameter count would have been the easy way from two Turing chips to one, but it costs capability. The team rebuilt the vision encoder instead, into HybridViT, a hierarchical version of TuringViT that passes 67% fewer image tokens to the language model and is trained against the large model’s outputs. That saving gives the language model room to keep most of its size on one chip. Owners of dual Orin-X cars in China were promised a solution later in 2026. About Europe, XPENG has said nothing.

The research behind it

XPENG publishes technical reports on the model work, which is unusual for a carmaker and useful for anyone trying to see where XPENG is heading. The framing comes from He Xiaopeng. Physical AI has four elements: models, computing power, data and physical embodiments, and the first three scale in software while the fourth is bounded by what can actually be manufactured and approved.

TuringViT is the vision encoder, the part that makes sense of what the cameras see. The XPENG Foundation Model Team published it on 23 June 2026. ViT stands for vision transformer, the model type nearly all image understanding in AI rests on. A model of that kind reads an image by comparing every part with every other part, a mechanism called attention. It is thorough but expensive: double the sharpness and the work quadruples, which is called quadratic growth. A car has several sharp cameras running at once and has to answer within a few milliseconds. XPENG’s answer is a block they call a Turing Block: five layers of Turing Linear Attention, or TLA, which scales close to linearly, and one layer of standard multi-head attention, which holds on to the fine detail TLA loses. The work then grows in step with the sharpness rather than with its square.

The more surprising claim is about the training. VISTA-Curation is the research team’s own way of sifting training data so that only the most instructive is left. By their account it matches the leading open-source encoders on about a tenth of the data. The same model is meant to serve assisted driving, the cabin and humanoid robots.

World models are the second track, and XPENG is explicit that they do not replace VLA. A VLA model learns from what humans did behind the wheel. That gives few examples, but each one is a whole decision. A world model learns from the footage itself: how cars, cyclists and traffic lights actually moved, frame by frame. That gives it a model of how traffic behaves, one that can be run forward to show what happens next. X-World writes new situations in seven camera angles from a text instruction and one real camera frame, so the car can be trained and tested on something that never happened. The research team splits it into three: X-Mind does the thinking, X-World makes the situations (control), and X-Foresight plays them forward (rollout). X-World is the workhorse of the three: by April 2026 it ran more than 500,000 simulation scenarios, up from 30,000 a year earlier, and XPENG puts its daily simulated mileage at the equivalent of 30 million real kilometres.

A world model running beside the VLA is too slow. The car would wait for the world model to compute a future, then wait again while the VLA decided from it. A car in traffic has no room for that wait. X-Mind, published 29 June 2026, is the answer: the prediction is folded inside the driving model itself, so it imagines how the situation will develop before it acts. It does not draw the future as pictures but as a sketch seen from above: twelve steps ahead, squeezed into 96 tokens, the units a model computes in. The sketch keeps only road topology, moving agents, traffic-light states, navigation intent and speed compliance. How much time those twelve steps cover, XPENG does not say. It comes out in one pass through the model rather than step by step, by a method they call recurrent block diffusion. All that engineering exists to make the thinking cheap enough to run in a car.

X-Foresight is the only one of the four names XPENG has attached to a car you can buy: the G9L in China. It appears in launch posts and on posters, not on the car’s equipment list. TuringViT, X-Mind and X-World exist so far only in the reports.

China and Europe are converging

The Chinese and the European cars come off the same engineering, which has never meant they could do the same things. The gap between a Chinese XPENG and a European one has been wide, and it is closing one difference at a time. I started following XPENG at the end of 2023, while I was weighing up a G9, and I drive one myself today. The gap has narrowed along the way. Two still separate them, and the first is settled the day the car is built.

Hardware. China sells chip count as a trim step, from one Turing to three, and prices it accordingly. Europe gets one chip on the P7+ and X9, two driving chips on a single L03 trim, and Bosch or Orin-X on everything else. That decides what a car can ever run, whatever arrives by OTA

Regulation. China and Europe do not run the same rules, and they are unlikely to. XPENG names UN Regulation No. 171 on Driver Control Assistance Systems as the route for NGP and VLA 2.0 in Europe. At its briefing on 27 August 2026 XPENG showed internal test footage under German, Dutch, French and Australian flags. Local acceptance testing of VLA 2.0 had already run in Germany, the company said. European approval is aimed at the first half of 2027. Once that lands, the heavy part is done: changes still need approving, but the distance to China gets shorter, not longer. The groundwork is older than the plan. XPENG joined the UN’s ADAS taskforce in 2023, and in February 2026 officials from the EU, the UK, the US, Japan and Canada rode through Shanghai in an XPENG running XNGP while the WP.29 automated-driving working group was meeting

LiDAR used to set the Chinese cars apart. The Chinese G9 of 2022 carried two units, and no export car has ever carried one. The cars on second-generation VLA carry none at all: neither the G9L, the GX nor the P7N runs LiDAR, and the Robotaxi programme is built on pure vision, without LiDAR and without HD maps.

The map has moved the same way. NGP leaned on map data at the start, and maps are not equally detailed everywhere. The car now reads mainly what the cameras see, so the quality of the map matters less for what it can do. The difference is still there, but it has moved from the driving to the screen: China runs Amap / AutoNavi (高德), because Google does not exist in the country, while the international cars have run TomTom and now get Google Maps on the L03, the G9L and later models.

Training data is the third difference closing. At Brand Day in Munich XPENG said VLA 2.0 accounted for 50.44% of driven kilometres on equipped cars after barely six months in China - more than half the driving done by the car itself. Those kilometres are Chinese, and European cars do not supply them in that volume. Two things pull the other way. First, the China-trained model performed close to home-market levels during local testing on German city roads with almost no local training data. Second, the X-World simulation engine generates new driving scenarios backed by a data throughput of over 100 million video clips per model training run, so a growing share of training was never a real drive.

The Chinese configurators split NGP into two line items, highway and city, and the Max trims get only the highway one. Under highway NGP sits a function on its way to the European cars: cabin vision and capacitive sensors notice a driver who has stopped responding, and the car pulls over and calls for help. It is on the L03 Ultra’s equipment list as part of VLA 2.0 NGP, so the first quarter of 2027 at the earliest.

The software reaches China first, and it reaches it more often. Staffing is one reason: at a meeting with XPENG’s ADAS team in Guangzhou in November 2025 I was told that at least ten times as many people work on the Chinese ADAS software as on the international one. Translation is another. An export build has to come out in every language the software carries and be adapted to each market in turn, work the home market never has to do. Then it has to be approved: when the 5.8.0 assistance regression needed a repair, the fix waited half a year on EU approval and reached European cars as 5.8.7 in March 2026. On the cockpit side China now runs XOS 5.x and 6.x in parallel while European customer cars stay on the 5.x train for now. The software is in Europe already, though: the L03 display cars run an unreleased build of XOS 6, and Google Maps appears to travel with it. The XOS version history lists every version.

XPENG Denmark put it plainly in a December 2024 xDIVE post: the European systems were not yet at the level of Guangzhou and Shanghai. Since then the G9L has launched as a global car with 64 markets named on day one.

Q&A

Sources

• XPENG’s XPILOT overview pages and the model comparison, retrieved 30 August 2026 - the 2.5 / 3.0 / 3.5 version numbers tied to the G3i, P7 Premium and P5, and the P7’s NVIDIA DRIVE Xavier
• CNBC, 24 October 2021 - XPILOT 3.5 launched for the P5
• XPENG online stores and price lists, Denmark, Germany and France, captured July 2026 - hardware lines, sensor counts and function lists per trim. The XPILOT label, and the L03 listing’s pairing of two Turing chips with VLA 2.0, re-checked on 29 August 2026 in the Danish, Dutch, Norwegian and Swedish stores
• XPENG owner’s manuals - P7+ MY26, New G9 MY25, G6 MY25, G9 MY23 and the X9 2026 international edition
• Release notes - XOS 5.9.5 for the custom APA hardware condition
• XPENG Brand Day, Munich, 16 July 2026 - Turing across cars, flying cars and robots; the VLA 2.0 share of kilometres driven in China; the European partner list
• XPENG press releases - the P7 launch of 27 April 2020 (first production car on NVIDIA DRIVE AGX Xavier), the G9 launch of 21 September 2022 with NVIDIA’s companion post (dual Orin-X, 508 TOPS, XNGP), the tape-out announcement of August 2024 and AI Day of November 2024 (the Turing chip’s 40 cores and 30-billion-parameter capacity)
• XPENG’s Turing chip pages - the English insight “One Chip, Triple the Power” and the Chinese chip page at xiaopeng.com - the NPU and ISP anatomy, the safety island and its certifications
• XPENG Denmark’s xDIVE series - the 17 December 2024 XPILOT hardware explainer: the sensor anatomy, and where the European systems stood against the Chinese cities
• XPENG newsroom - AI Day, 5 November 2025 (the 72-billion-parameter cloud model; Volkswagen as launch customer), the WP.29 demonstration post of 13 February 2026, and the X-World technical-report release of 29 April 2026
• XPENG launch material for the GX and the G9L, August 2026, with the company’s Weibo and @XPENG_Global posts of 26-27 August - chip counts per trim, VLA 6.3.0, Infini-VLA, X-Foresight, Master Agent and the UN R171 approval route
• XPENG official corporate release on second-generation VLA, 27 August 2026 (published 1 September 2026) - the 4D spatiotemporal model, Infini-VLA (30 s), X-Foresight (6 s), Streaming Inference (+300%), Master Agent, Turing VLA 2.0 Lite on G9L Max, local validation testing in Germany and the IRON robot’s 3 Turing chips (2,250 TOPS)
• XPENG’s Weibo post of 29 August 2026 with the clip of He Xiaopeng’s talk - second-generation VLA Lite: the HybridViT slide with the 67% cut in image tokens, the rollout from September and the promise to dual Orin-X owners. The post’s own text mentions only the distillation and September; the rest is from the clip
• XPENG Foundation Model Team technical reports - TuringViT (23 June 2026) and X-Mind (29 June 2026), with Xianming Liu’s world-model posts of 4 June 2026, his CVPR workshop talk of 29 June 2026 (the split into controllable generation, long-horizon forecasting and proactive reasoning) and He Xiaopeng’s Physical AI essay of 5 July 2026
• Meeting with XPENG’s ADAS team in Guangzhou around AI Day, 5 November 2025 - the headcount on the Chinese ADAS software against the international one, told on site. The trip is written up in Visiting XPENG
• NVIDIA DRIVE AGX Orin and NVIDIA’s DRIVE AGX developer documentation - the 254 TOPS figure, the INT8 basis and the ISO 26262 safety island. XPENG has published no compute figure for these cars

Last reviewed: August 2026