Realtime AI News
StepFun unveils STEPX Neo, a native AI agent phone entering production
StepFun (阶跃星辰) has unveiled the STEPX Neo, a native AI agent smartphone that integrates large language model capabilities at the operating system level. The device has entered the production scheduling phase, marking a milestone for AI-native hardware in China.
StepFun, the Chinese AI company known for its STEP series of large language models, has officially unveiled the STEPX Neo — a smartphone built from the ground up with native AI agent capabilities. The product has already entered the production scheduling phase, according to a report from NetEase mobile channel.
The STEPX Neo is not a conventional smartphone with AI apps bolted on. Instead, it integrates large model capabilities directly into the operating system layer, making AI agent functionality a native component of the device rather than an add-on feature.
Entering the production scheduling phase is a significant milestone, indicating that the device has passed prototyping and validation and is moving toward mass production. This brings the product from concept to reality.
StepFun is best known for its family of STEP foundation models. The move into hardware represents a strategic expansion from the model layer into applications and devices, positioning the company to compete in the emerging AI-native device category.
AI-native phones are seen by many as the next frontier beyond traditional smartphones. By embedding AI at the system level, these devices promise deeper, more natural agent interactions — the phone understands context, anticipates needs, and executes tasks proactively.
The STEPX Neo launch comes at a time when multiple Chinese tech companies are racing to define the AI hardware form factor. StepFun's choice of the smartphone form factor gives it the advantage of an existing user paradigm, but also pits it against established smartphone giants.
Key questions ahead include the STEPX Neo's final specifications, pricing strategy, and availability timeline. The real-world performance of its AI agent capabilities — particularly response speed and accuracy in offline scenarios — will determine its market reception.
Why it matters
StepFun's STEPX Neo represents a strategic leap from AI model provider to AI hardware maker, adding significant momentum to the AI-native device category in China.
Nearby Updates
All07/17, 15:56
Midu Launches DataQ Data Analysis and Insights Agent at WAIC 2026
Midu has unveiled DataQ, a data analysis and insights intelligent agent, at the 2026 World Artificial Intelligence Conference (WAIC) in Shanghai. The agent enables users to query data, detect trends, and generate reports through natural language interaction, lowering the barrier to professional data analysis.
07/17, 16:09
Alibaba's Qoder Tops China AI Coding Market with 47.6% Share, Exceeding #2-5 Combined
IDC's latest report shows Alibaba's Qoder captured 47.6% of China's AI coding market revenue in 2025, exceeding the combined share of the second through fifth competitors. The agent-powered coding platform, launched less than a year ago, has reached 5 million users by offering a comprehensive product line spanning IDE, cloud agent, CLI, and digital employee tools.
07/17, 18:00
OpenAI CFO Introduces AI Scorecard to Help Businesses Quantify AI ROI
OpenAI CFO Sarah Friar has introduced a practical AI Scorecard framework designed to help enterprises measure the real return on their AI investments. The framework evaluates AI performance across four dimensions: useful work, cost per successful task, dependability, and return on compute.
07/17, 18:11
WeRide Launches Physical AI Foundation Model WITT, Processing 10,000 Minutes of Video Per GPU Per Day
WeRide has unveiled WITT (World Intelligence Toward Truth), a physical AI foundation model that introduces the concept of "minimum physical fact units" to transform autonomous driving video into structured, searchable data. The company claims WITT can boost data processing efficiency by up to 200x, handling 10,000 minutes of driving video per day on a single GPU at one-third the factual error rate of general-purpose large models.