DeepRoute IO by DeepRoute.ai: status and timeline

Milestones

Next · One million vehicles and 10+ models targeted

DeepRoute IO first deployed on production vehiclesSep 2023Complete.
Map-free autonomous driving debuts in ChinaSep 2023Complete.
First 20,000 vehicles delivered with DeepRoute IO (August 2024)Aug 2024Complete.
200K vehicles and 40% market share by Oct 2025Oct 2025Complete.
Series C1 funding closes at $100 million from Great Wall Motor5 Nov 2024Complete.
DeepRoute IO 2.0 and 40B VLA model launchAug 2025Complete.
One million vehicles and 10+ models targetedTarget 2026Current milestone.
International OEM partnerships and regulatory approvals plannedTarget 2027Not yet reached.

Most important updates

Upcoming

2026One million vehicles and 10+ models targeted (next)
Q4 2026International OEM partnerships announced
2027International OEM partnerships and regulatory approvals planned

Current obstacles

Rare events and edge casesModel fails on unusual situations: night jaywalking, road debris, extreme weather. Validating safety for all cases is hard.
International regulationChina's approval is faster than US and Europe. Each market has different safety standards and liability rules.
Model confidence and fallbackLarge models sometimes output plausible but wrong answers. Safety requires detecting low-confidence decisions and falling back to human control.

Physics limits

Data collection for new marketsModels trained on Chinese highways must be retrained for US or European roads: different road markings, traffic patterns, weather. Collecting millions of miles takes years.
Real-time inference latencyDecisions must happen in under 100 milliseconds or the car is unresponsive. Larger, more accurate models take longer to run; smaller models are less accurate. This trade-off is hard.
Camera limitations in darkness and rainHeavy rain, snow and night scenes are harder for cameras to perceive than human eyes. Lidar sees better but costs 5-10x more. This limits where camera-only driving works.

How it works

4 parts

Input

Camera feeds

Multiple cameras capture the road, other vehicles, pedestrians and traffic signals. No maps needed.

Reasoning

40-billion parameter model

Processes images to understand the scene: where is the car, what are others doing, what are traffic rules. Outputs a plan for the next few seconds.

Control

Steering and acceleration

Translates the plan into steering, throttle and brake commands. A safety layer filters unsafe decisions before they reach the vehicle.

Deployment

Mass-market vehicles

Runs on NVIDIA DRIVE AGX Thor (Blackwell GPU). Integrated into Great Wall Motor, Geely and Smart models for SAE Level 2-3 driving assistance.

Spec sheet

Spec	DeepRoute IO
Vehicles delivered by end 2025	200,000+^R (reported)
Urban NOA market share (Oct 2025)	40%^R (reported)
Vehicle models supported	15+^R (reported)
IO 2.0 VLA model size	40 billion parameters^R (reported)
Primary compute platform	NVIDIA DRIVE AGX Thor (Blackwell GPU)^R (reported)

^R reported by the company

Papers & demos

3 items

Mar 2026demo
DeepRoute at GTC 2026: unified VLA cuts training cycles 5 days → 12 hours (opens prnewswire.com)Unified model (driver/analyst/critic) automates fleet learning; targets one million vehicles by end 2026.
Aug 2025demo
DeepRoute IO 2.0 announced; 40B VLA model in production vehicles (opens prnewswire.com)First mass-production VLA model in consumer vehicles; no HD maps required.
Sep 2023demo
DeepRoute IO launches with map-free autonomous driving (opens automotiveworld.com)First deployment of map-free navigation in China; validated real-world viability of perception-only architecture at scale.

Update log

3 updates

About DeepRoute.ai

The team behind DeepRoute IO

DeepRoute.ai

Urban autonomous driving, VLA large models, mass production

DeepRoute.ai develops autonomous driving software for Chinese OEMs using large vision-language-action models that reason about driving decisions from camera input. Founded February 2019 in Shenzhen, now deployed in 200,000+ vehicles.

Deployed 200,000 vehicles by end 2025; went from zero in early 2024 to 40% market share in China.
Series C1 of $100M from Great Wall Motor (Nov 2024) funds VLA model development and global partnerships.
IO 2.0 (Aug 2025) runs 40B-parameter VLA model without maps; NVIDIA DRIVE AGX Thor compute platform.

deeproute.ai (opens deeproute.ai)Company profile →

Founded: 20197 yrs
Headquarters: China
Status: Private
Valuation: private
Raised: $100M1 round
Last round: Series C1 · $100M5 Nov 2024
Works in: VehiclesSelf-driving software
Coverage: 1 program · 4 updateslatest 1 Mar 2026checked 10 Oct
Lead investors: Great Wall Motor
Partners: Geely (opens techcrunch.com)Smart (opens techcrunch.com)NVIDIA (opens autonews.gasgoo.com)
Investors: Great Wall Motor (opens therobotreport.com)
People: Zhou Guang (opens en.wikipedia.org)CEO & Co-founder

DeepRoute IO

Milestones

Most important updates

Upcoming

Current obstacles

Physics limits

How it works

Camera feeds

40-billion parameter model

Steering and acceleration

Mass-market vehicles

Spec sheet

Papers & demos

Update log

Mar 2026

Oct 2025

Aug 2025

About DeepRoute.ai

Current obstacles

Physics limits

How it works

Camera feeds

40-billion parameter model

Steering and acceleration

Mass-market vehicles

Spec sheet

Papers & demos

Update log

Mar 2026

Oct 2025

Aug 2025

About DeepRoute.ai