StepX bets on a model-native agent phone with its first device, Neo
What's the deal? Chinese company StepXDealroom has a profile for this one. Try Dealroom → released its first large-model-native agent phone, STEPX Neo, on October 13. Rather than bolting AI onto an existing handset, it designed the model, system, and hardware around what a personal agent needs.
Why now? Personal agents are becoming the AI industry's main storyline. GoogleDealroom has a profile for this one. Try Dealroom → launched Gemini Spark, Meta released Muse, and OpenAI entered with its always-on Dots agent.
Investors are backing the field heavily: personal AI assistant company Town raised $55 million in a Series A, Hark secured $700 million, and Instinct reportedly raised $250 million at a $2.5 billion valuation.
What's the endgame? StepX is building what it calls a "software-hardware-model" loop. The phone collects real-world context, the system links it to memory, tools, and permissions, and the model identifies intent and plans tasks.
Most AI phones start from an existing device, StepX argues, adding search, photo editing, or cross-app features piecemeal. StepX instead defines the agent's needs first, then designs everything else around them — comparing rivals to installing smart furniture in an old house while it builds a new one.
How does it work? The model layer runs a matrix of Pro, Flash, and Edge models: cloud-based Pro handles complex tasks, Flash balances quality and speed, and on-device Edge handles instant or privacy-sensitive work. A routing system keeps local tasks on the device and sends deep reasoning to the cloud.
Its Step AOS operating system treats the agent as a system-level capability, with a "dual-entry" design: a familiar app desktop alongside the Amoo agent space. The system also breaks phone functions into atomic capabilities and supports A2A, MCP, Skill, API, and CLI protocols to connect external apps and agents.
StepX also launched Step Studio, a development tool that lets users build AI apps, games, or new agents through natural language. A "dual-domain, six-layer" memory system stores user facts and agent knowledge, aiming to recall useful information at the right moment rather than simply store more data.
What could go wrong? The native route is heavier and harder than feature integration. Whether the full chain — from understanding a user to calling services and completing tasks — runs reliably remains the real dividing line.
The signal: The release lands amid a crowded Chinese AI phone push, with Doubao's second-generation device and handsets from HuaweiDealroom has a profile for this one. Try Dealroom →, Xiaomi, OPPO, and vivo all arriving. The contest is no longer about chat tools, but over who owns the next personal entry point — a device that understands users over time and acts on their behalf.
Read more: 36kr.com
Image credit: Generated with Gemini
