Doubao Phone Released: Inside the World's First AI Agent Phone

September 16, 2026 — At 2 p.m. today, ZTE's nubia and ByteDance officially launched the nubia NaviX Ultra, widely known as the “Doubao phone.” It is being billed as the world's first AI agent phone — and the first AI agent phone to clear both model registration and terminal network licensing. Here is what it is, how the industry got here, and where AI phones go next.

Doubao phone at a glance

Specs above are compiled from official teasers and media reports; treat the final launch page as the source of truth.

An AI agent phone is not another “AI phone”

For two years, “AI phone” has been stretched thin: AI retouching, AI summaries, a rebranded voice assistant. Those products share one architecture — the operating system is the core, and AI is an app running on top of it.

The NaviX Ultra flips that structure. The large-model agent sits underneath the operating system as its scheduling engine, present in system calls, app interactions, and permission management. You are not downloading an assistant; the assistant is the phone.

nubia summarizes the pitch in four verbs: hear, act, remember, protect.

• Hear. Dialect support, noise-robust wake-up, and an AI button with fingerprint authentication — verified once, and the assistant works even from the lock screen.
• Act. The “operate the phone” feature ships in beta: the agent can tap, scroll, and chain steps across apps. Official demos include pointing the camera at a room and asking for a bookshelf under one meter and ¥1,000, or saying “get me a ride to the office” and watching it confirm the address, pick a car type, and place the order. Payments and other high-risk steps still require manual confirmation.
• Remember. With permission, the assistant searches local photos, messages, and notes — ask when your travel permit expires and it finds the answer on your phone. Meeting audio gets transcription, translation, speaker labels, and follow-up questions via Feishu Minutes.
• Protect. Minimum-necessary data collection, user-controlled authorization, tiered agent permissions, and a full audit trail for cross-app operations.

Fifteen years in three acts: how AI phones got here

Act I (2011–2018): AI as a feature, not a capability

Siri arrived on the iPhone 4s in 2011. Honor's Magic Live system followed in December 2016 with semantic understanding, computer vision, and on-device decision making — arguably the first true prototype of an AI phone. Then came 2017, the “AI chip year”: Apple's A11 and Huawei's Kirin 970 both shipped dedicated NPUs.

And then the hype cooled. The hardware existed, but the killer app did not. NPUs spent their cycles on face unlock and camera tuning, users barely noticed, and the promised upgrade wave never came. Lesson one: compute without a model to run on it is dead weight.

Act II (2023–2025): large models move in, then hit the wall

The generative AI wave changed the equation. Huawei's Mate 60 shipped with Pangu in August 2023; vivo's X100 carried a 7-billion-parameter on-device model that November; OPPO's Find X7 and Xiaomi's 14 Ultra followed. 2024 became the industry's declared “Year One of the AI phone.” By 2025, on-device models were table stakes on flagships.

Then came the pivotal experiment. On December 1, 2025, ByteDance and ZTE's nubia released the first Doubao phone — the M153, a 3,000-unit engineering preview priced at 3,499 RMB. It sold out instantly; resale prices hit 36,000 RMB and rental services charged 50–79 RMB per day. Within a week, reality landed: WeChat flagged login anomalies, Taobao threw up human-verification walls, and major banking apps blocked the device outright. The first generation relied on GUI-level simulated taps — effectively reaching into apps the way a person would — and the super-apps shut the door. Financial-app operations were disabled by December 5.

Lesson two: a technology that outruns its ecosystem crashes into the ecosystem wall.

Act III (2026): the year of the agent

• June 18, 2026 — eight Chinese ministries publish the implementation opinion on accelerating “AI + consumption,” elevating AI devices to a national strategy.
• July 15, 2026 — China's Cyberspace Administration publishes the first centralized filing of on-device generative AI services covering seven vendors: Apple, Huawei, OPPO, vivo, Xiaomi, Samsung, and nubia. When a regulator formally names a technology, it has stopped being a lab demo.
• September 1, 2026 — the NaviX Ultra receives its network license, completing the full compliance chain from model registration to terminal certification.
• September 14, 2026 — ByteDance releases the consumer edition of Doubao Phone Assistant; pre-orders open and approach 270,000 within a day, then pass 360,000.
• September 16, 2026 — today's launch.

Why this launch is different: three shifts

1. From simulated taps to protocol cooperation

Generation one operated apps by imitating a human. Generation two switches to MCP (Model Context Protocol) and A2A (Agent-to-Agent protocol): apps open interfaces to the agent instead of being puppeted from outside. The accompanying SAEP declaration protocol lets any third-party app declare its own boundaries — allowing or refusing AI automation inside it — with a 30-day public comment period. The choice is handed back to the apps. JD-backed Doubao services already connect to Didi, Luckin, Qunar, and Caocao mobility.

Protocol adoption is necessary but not sufficient, as East China Normal University professor Xiangfeng Wang put it: the first generation was blocked because the ecosystem would not cede control, and a protocol by itself does not settle whether an app consents at all.

2. From cloud to device

On-device inference attacks three problems at once: latency (local responses in milliseconds), privacy (sensitive data never leaves the phone), and cost (light tasks stop burning cloud compute). The hardware finally supports it: flagship NPUs now exceed 70 TOPS and can run quantized 13-billion-parameter multimodal models, with a 7B model compressed to roughly 3.9 GB of memory.

3. From engineering preview to mass-market product

This is the quiet but heaviest milestone. The M153 was a 3,000-unit experiment; the NaviX Ultra is a fully certified flagship with more than 360,000 pre-orders — twelve times the first run. The AI agent phone has moved from proof of concept to commercial scale.

What is still unsolved

An honest read includes the open problems:

• The ecosystem wall stands. Protocols solve “how to connect,” not “whether to connect.” Super-apps make money on ads and recommendation feeds, and an agent that skips both undermines that model.
• Liability is undefined. When an agent acts on your behalf and gets it wrong, who answers? The industry's current answer — force manual confirmation for payments — is a stopgap, not a framework. Lawyers and analysts are already calling for regulator-led rules instead of platform-written ones.
• Trust takes time. Generation one's privacy controversies and the disabled financial features left a mark. Omdia analyst Zhong Xiaolei's framing is the right one: AI phone competition is system versus system, not brand versus brand.

Where AI phones go next: five bets

1. App icons give way to intent

The three-layer structure we have built over fifteen years — open the app, find the feature, tap the button — compresses into a sentence. When users no longer need to know which app owns a function, the value anchor of the operating system shifts from the app store to intent routing. Whoever owns the intent owns the entry point.

2. Long-term memory becomes the real moat

Benchmarks can be matched within a generation. A personal context that has absorbed three years of your preferences cannot. That is why every vendor is racing to build local memory search. The phone stops being a device and starts being a continuously evolving digital counterpart — one that remembers your permit expiry dates, your phrasing habits, the proposal you never finished.

3. Agents collaborate across devices

A2A is not just about one phone. When agents on your phone, PC, car, watch, and glasses can talk to each other, personal AI stops being tied to a single screen. The phone likely becomes the portable compute and identity hub while the interface spreads across earbuds, glasses, and pendants.

4. The business model gets rewritten

If AI completes the purchase, the booking, and the order while skipping ads and feeds, the revenue structure between brands, apps, and device makers has to be renegotiated: hardware plus AI subscriptions, intent-distribution revenue shares, and visibility competition inside AI recommendations. The next battleground of SEO may be whether the assistant recommends you at all.

5. Form factors split — and small wins

Not everyone wants a six-inch flagship loaded with everything. The market is splitting between all-powerful agent hubs and deliberately light, pocket-sized AI companions — small screens, card-format devices, AI pendants, keyboard-equipped sidekicks that chase no benchmarks and simply get things done from a sentence. Mass adoption rarely comes from the most powerful device; it comes from the one that is actually carried and actually used. That is the bet behind compact AI phones like the Meydo C1: a 4-inch AI companion built to live in your pocket rather than dominate your hand.

FAQ

Who makes the Doubao phone?

It is not a ByteDance-built handset. ZTE's nubia brand handles hardware R&D and manufacturing; ByteDance supplies the Doubao large model. The product is the nubia NaviX Ultra.

How much does it cost?

The first-generation M153 launched at 3,499 RMB as a limited engineering preview. The second-generation NaviX Ultra is reported at 4,999 RMB; treat the official launch page as final.

How is it different from a regular AI phone?

Regular AI phones treat AI as an app-layer feature. The Doubao phone embeds the model agent into the OS itself, enabling cross-app task execution, local memory, and on-device inference. One is a phone with AI features; the other is AI-native.

Can it operate WeChat?

Not the way generation one tried. The first model simulated taps and was blocked by WeChat and other super-apps. Generation two only connects where an app deliberately exposes MCP services and grants permission, and apps choose whether to allow automation.

What is an AI agent phone?

A phone built around a large-model agent as the core of the operating system. Interaction shifts from command-driven to intent-driven: you state a goal, the agent plans and executes across apps, and hands back control for high-risk steps. We cover the fundamentals in What Is an AI Agent Phone?

Are AI phones selling well?

Yes. Aggregated analyst estimates put China's 2026 AI phone shipments near 147 million units, up 31.6% year over year, with penetration crossing 50% for the first time; global generative-AI phone penetration is expected to pass half of the market by 2027.

The takeaway

The NaviX Ultra matters less as a new handset than as a settled consensus: phone competition is moving from “whose hardware is better” to “whose AI gets more done.” In 2017, the industry held NPUs it did not know how to use. In 2026, the answer has arrived: use them to work on your behalf.

What happens next depends on two things — whether the ecosystem opens its doors, and whether users are willing to hand over the keys. The moment both happen, the next real upgrade cycle begins.