DEPLOYDatabase

Brain

XPENG VLA 2.0

XPENG's mass-produced billion-parameter-scale foundation model for physical AI. Uses a novel 'Vision-Implicit Token-Action' path that eliminates the language translation step, achieving end-to-end direct generation from visual signals to action commands. First in industry to deploy billion-parameter model on-vehicle (2250 TOPS via Turing AI chip).

Cross-domain: simultaneously drives AI cars, IRON humanoid robot, and flying cars. Trained on ~100M real driving video clips without annotation. Rolled out to XPENG Ultra models Q1 2026.

Foundation model · Maturity: Production · Closed

Machine-readable surfaces

Architecture

Vision-Implicit Token-Action (V-I-A), 72B cloud base + billion-parameter on-vehicle, Turing AI chip 2250 TOPS

Key facts

VLA 2.0 architecture and on-vehicle scale (primary XPENG)
XPENG's Nov 5 2025 AI Day release introduces VLA 2.0 with a Vision-Implicit Token-Action path that removes a language translation step for end-to-end vision-to-action generation. It cites a 72 billion parameter cloud base model iterating about every five days on a 30,000-card cluster, and says a billions-of-parameters VLA 2.0 build was installed on Ultra models with 2250 TOPS via Turing AI chip software-stack work. Treat scale and architecture claims as company-stated (XPENG)
Rollout and Volkswagen (primary XPENG)
The same release says pioneer users would experience VLA 2.0 by end of December 2025 with a full Ultra-model rollout in Q1 2026, that VLA 2.0 would be open-sourced to business partners, and that Volkswagen is the launch customer for VLA 2.0 while the Turing AI chip secured a Volkswagen nomination. Treat partner and schedule claims as company-stated (XPENG)
VLA 2.0 Physical AI unveil (primary XPENG)
XPENG's pressroom piece titled on Physical AI Emergence unveils XPENG VLA 2.0 alongside Robotaxi, next-gen IRON, and flying systems, describing VLA 2.0 as a core Physical AI application for vehicle control. Treat product framing as company-stated (XPENG)
End-to-end physical AI model
XPENG describes VLA 2.0 as an end-to-end physical AI model that maps visual signals to action commands for AI-defined vehicles, humanoid robots, and flying vehicles (XPENG VLA 2.0).
Autonomous-driving deployment
XPENG says VLA 2.0 supports full-scenario driving functions and is being advanced through public-road testing and a planned delivery programme (XPENG VLA 2.0 deployment).
Physical AI world model
XPENG's CVPR 2026 pressroom release presents its Physical AI and world-model work as part of the company's autonomous-driving technology roadmap (XPENG at CVPR 2026).
X-Mind future forecasting
XPENG describes X-Mind as a future-foresight brain that uses visual chain-of-thought and predictive world-model reasoning to forecast future traffic states before driving decisions; treat the architecture framing as company-stated (XPENG X-Mind).
Vision-language-action driving model
XPENG describes VLA 2.0 as a physical AI model that translates visual signals into action commands and supports assisted and autonomous driving systems (XPENG VLA 2.0).
Global deployment plan
XPENG says it is expanding public-road testing of VLA 2.0 and plans global vehicle delivery using the model in 2027 (XPENG deployment plan).

Developed by (1)

Common questions

What is XPENG VLA 2.0?
XPENG's mass-produced billion-parameter-scale foundation model for physical AI. Uses a novel 'Vision-Implicit Token-Action' path that eliminates the language translation step, achieving end-to-end direct generation from visual signals to action commands. First in industry to deploy billion-parameter model on-vehicle (2250 TOPS via Turing AI chip). Cross-domain: simultaneously drives AI cars, IRON humanoid robot, and flying cars. Trained on ~100M real driving video clips without annotation. Rolled out to XPENG Ultra models Q1 2026.
Who developed XPENG VLA 2.0?
XPENG VLA 2.0 is credited to XPeng on the DEPLOY registry. Each developer attribution is verified via primary sources.
Is XPENG VLA 2.0 open source?
No. XPENG VLA 2.0 is recorded as proprietary on the DEPLOY registry. Model weights and source are not publicly available.
What type of AI is XPENG VLA 2.0?
XPENG VLA 2.0 is a foundation model, built on a Vision-Implicit Token-Action (V-I-A), 72B cloud base + billion-parameter on-vehicle, Turing AI chip 2250 TOPS architecture on the DEPLOY registry.
What is XPENG VLA 2.0's maturity stage?
XPENG VLA 2.0 is at the production stage on the DEPLOY maturity ladder. Production stage means high-volume sustained commercial operations.
Which robots run on XPENG VLA 2.0?
No robot models on the DEPLOY registry are recorded as running XPENG VLA 2.0. DEPLOY wires brain-to-model connections only when the wiring is verifiable from primary sources; absence may reflect pre-deployment or unverified manufacturer claims.

Sources (5)

  1. https://www.xpeng.com/pressroom/news/019a56f54fe99a2a0a8d8a0282e402b7
  2. https://www.xpeng.com/news/019a56f54fe99a2a0a8d8a0282e402b7
  3. https://www.xpeng.com/pressroom/news/019cae5e67b99c0960ee8a028129016a
  4. https://www.xpeng.com/pressroom/news/019e95aed2899e8226228a028d650112
  5. https://www.xpeng.com/pressroom/news/019f12539bff9f1220b48a028223000e
Methodology: Verified · 5 sources (no primary) · last reviewed 2026-10-09

Verification posture

Verified

Low confidence

Review state

Stable

Last reviewed 2026-10-09

Maturity + lifecycle

Maturity stage: production

Sources by quality tier

5
unclassified
Unclassified source

The framework is documented at /methodology. Corrections at /corrections. Reviewer: DEPLOY editorial team.

Methodology surface for XPENG VLA 2.0.

Canonical ID 61c27e6d-6315-44b9-9ebf-128b8c6a68de