Qualcomm and partners deploy 30-billion-parameter AI model on smartphones for on-device agents
At the 2026 Snapdragon Summit, Qualcomm Technologies partnered with StepFun, Wulianghuo, and Longsys to deploy the StepEdge-Omni 30B-MoE AI model on the 6th Gen Snapdragon 8 Super Elite mobile platform. The collaboration demonstrated an on-device AI agent capable of email comprehension, itinerary planning, and calendar syncing without cloud dependency. Through heterogeneous scheduling and storage optimization, runtime memory requirements were reduced by over 50%, achieving prefill throughput exceeding 330 tokens/s and decode throughput exceeding 28 tokens/s.
IllustrationEditorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Common ground
- Edge AI is real and important, especially for privacy, low latency, and offline use in places with poor internet.
- Running a 30-billion-parameter model on a smartphone is a genuine technical achievement.
- The Global South and developing nations benefit from on-device AI that doesn't need constant cloud access.
- Qualcomm is making a strategic pivot to stay relevant as its core mobile chip market matures.
Points of contention
- Whether this partnership is a geopolitical revolution or just a smart business deal.
- Whether China is leading edge AI or catching up to work already done by Apple, Google, and Samsung.
- Whether running AI on American-designed chips means China is still dependent or building true sovereignty.
- Whether edge AI can replace cloud AI for complex reasoning tasks or is only good for simple, everyday uses.
Blind spots
- Both sides downplay the critical dependency on TSMC's 2nm fab in Taiwan, which neither the US nor China fully controls.
- The debate ignores how export controls are actually speeding up China's domestic chip development, not just slowing it.
- Neither side fully addresses the power and battery drain of running sophisticated AI agents continuously on a phone.
WorldAttention’s read
This debate shows a clear split: one side sees a practical business deal where Qualcomm sells chips and Chinese partners optimize software, while the other sees a shift toward a decentralized, privacy-first AI world led by Chinese innovation. Both agree edge AI matters and the 30B model on a phone is impressive, but they clash on whether this is a geopolitical breakthrough or just smart marketing. The real blind spot is that everyone still depends on TSMC's advanced fabs in Taiwan, and no one has a solid answer for when that changes. In the end, edge AI is a real and growing trend, but it won't replace cloud computing for heavy lifting—it's a tool for specific tasks, not a total revolution.
Reporting timeline
Qualcomm, StepFun, and Longsys Partner to Run 30B AI Model on Smartphones
During the 2026 Snapdragon Summit on September 23, Qualcomm Technologies announced a partnership with StepFun, Wulianghuo, and Longsys (09976) to adapt and optimize the StepEdge-Omni 30B-MoE (Mixture of Experts) model for on-device inference on the 6th Gen Snapdragon 8 Super Premium mobile platform. A live demonstration showed an on-device AI agent capable of tasks like email comprehension, itinerary planning, calendar syncing, and flight/hotel recommendations without cloud calls. Through collaborative optimization of the inference engine, CPU/GPU/NPU heterogeneous scheduling, and storage, the model's runtime memory demand was reduced by over 50%, achieving a prefill throughput exceeding 330 Token/s and a decode throughput exceeding 28 Token/s. Longsys contributed its end-side storage capabilities, enabling model weights to be loaded from flash memory to reduce persistent RAM usage. Analysts noted that as AI phones evolve from simple Q&A to autonomous planning agents, demand for high-speed, high-capacity, low-power storage will increase, deepening ecosystem collaboration among storage, chip, model, and device makers.
Read sourceQualcomm, StepFun, and Longsys Partner to Run 30B AI Model on Smartphones
On September 23, during the 2026 Snapdragon Summit, Qualcomm Technologies announced a partnership with StepFun, Wulianghuo, and Longsys (stock code: 09976) to adapt and optimize the StepEdge-Omni 30B-MoE (Mixture of Experts) model for on-device deployment on the sixth-generation Snapdragon 8 Super Premium mobile platform. The collaboration demonstrated an on-device AI agent capable of tasks such as email comprehension, itinerary planning, calendar synchronization, flight and hotel recommendations, and email drafting without cloud dependency. Through coordinated optimization of the inference engine, CPU/GPU/NPU heterogeneous scheduling, and storage solutions, the model's runtime memory requirements were reduced by over 50%, achieving prefill throughput exceeding 330 tokens per second and decode throughput exceeding 28 tokens per second in live tests. Longsys contributed end-to-end storage capabilities, enabling model weights to be loaded from flash memory to reduce persistent RAM usage. Analysts cited in the article note that as AI phones evolve from simple Q&A to autonomous planning and continuous task execution, demand for high-speed, high-capacity, low-power storage will increase, deepening ecosystem collaboration among storage, chip, model, and device manufacturers.
Read sourceQualcomm unveils 2nm chips, runs 300B-parameter AI model locally on phone
At the 2026 Snapdragon Summit on September 22, Qualcomm launched the sixth-generation Snapdragon 8 Super Elite and Snapdragon 8 Elite mobile platforms, both built on a 2nm process, breaking its tradition of releasing a single flagship SoC. The chips will be used by Xiaomi, vivo, OPPO, and Honor. Qualcomm CEO Cristiano Amon emphasized that the 'agent era has arrived,' with AI evolving from a tool that responds to commands into an agent that understands context, decomposes tasks, and executes actions with authorization. To enable complex on-device AI agents, Qualcomm partnered with StepFun, Wulianghuo, and Jiangbo Long to deploy a 300-billion-parameter MoE model (StepEdge-Omni 30B-MoE) locally on the Snapdragon 8 Super Elite. This is notable because most on-device models are in the single-digit-billion range. The MoE architecture allows the model to activate only a subset of experts per inference, reducing memory and compute demands. Wulianghuo reduced memory requirements by over 50% through heterogeneous scheduling across CPU, GPU, and NPU, while Jiangbo Long optimized storage-compute coordination. The chip features a custom Oryon CPU reaching 5GHz, a GPU with dedicated AI cores, and an NPU with a Transformer accelerator. Qualcomm also plans to bring its HBC (high-bandwidth computing) technology from data centers to Snapdragon platforms. The article notes that mobile chip competition is shifting from raw performance to how large an AI model can run locally and how much it can accomplish for users.
Read sourceShow 3 older updatesHide older updates
Qualcomm Snapdragon Summit Showcases On-Device AI Agents, Shifting Focus to Local Intelligence
The 2026 Snapdragon Summit highlighted a shift in mobile computing, with Qualcomm demonstrating on-device AI agents capable of complex tasks like email understanding, calendar syncing, and trip planning without cloud reliance. Qualcomm's VP of Communications, Clare Conley, noted in a bylined article that the focus is moving from technical specs to practical user benefits. CEO Cristiano Amon described the ideal agent experience as personalized, proactive, responsive, trustworthy, and effortless. To enable this, Qualcomm launched the Snapdragon 8 Super Elite and Snapdragon 8 Elite platforms, featuring the Oryon CPU, Adreno GPU, and Hexagon NPU with new capabilities for Transformer workloads. A key partnership with Step, Wulianghuo, and Jiangbo Long successfully deployed a 30-billion-parameter MoE model on-device. The summit emphasized that no single company can build the agentic future alone, highlighting collaborations with Mastercard for payments and multiple OEMs like Xiaomi, vivo, and OPPO. The article concludes that this era is redefining the human-technology relationship from tool interaction to system collaboration.
Read sourceQualcomm, Partners Deploy 30-Billion-Parameter MoE AI Model on Smartphone for On-Device Agents
During the 2026 Snapdragon Summit on September 22, Qualcomm announced a four-party partnership with StepFun, Wulianghuo, and Longsys to deploy the StepEdge-Omni 30B-MoE model on the 6th Gen Snapdragon 8 Super Elite mobile platform. The collaboration aims to create an on-device AI agent assistant with autonomous service and personalization capabilities. The MoE (Mixture of Experts) model architecture activates only necessary expert modules per task, reducing computational load and memory bandwidth compared to dense models. The new platform features Qualcomm Oryon CPU (up to 5GHz), Adreno GPU with Matrix Cores, and Hexagon NPU with Element Accelerator. Wulianghuo's Ember inference engine reduced memory requirements by over 50% without model modification or accuracy loss. Longsys provided on-device AI storage solutions for efficient model weight loading. The demonstration showed a complete autonomous workflow from email understanding to travel planning, calendar syncing, and email drafting, all processed locally. Performance metrics include prefill throughput exceeding 330 tokens/s and decode throughput exceeding 28 tokens/s. Qualcomm's Chris Patrick stated the results enable fast, privacy-focused AI experiences on smartphones. StepFun's Yu Gang noted the platform's on-device computing power unlocks low-latency, high-privacy applications. Wulianghuo's Wang Yukun emphasized the memory reduction allows 30B models to run on phones. Longsys's Huang Qiang highlighted the importance of storage-compute synergy for scalable on-device AI.
Qualcomm's Snapdragon Summit: Agentic AI Era Requires Ecosystem Collaboration
This article from 通信世界网 analyzes Qualcomm's vision for the 'agentic AI era,' as presented at the Snapdragon Summit by CEO Cristiano Amon and CMO Don McGuire. The central thesis is that AI is evolving from a simple tool that responds to commands into a collaborative partner that understands context, takes action, and helps users complete tasks across multiple devices. Amon emphasized that the agentic AI experience must be personalized, proactive, responsive, trustworthy, and effortless, requiring AI to run on edge devices like smartphones, PCs, cars, and wearables. The article highlights that no single company can build this future alone; collaboration across the ecosystem is essential. Qualcomm positions its Snapdragon platform as the processor for this new era, focusing on on-device inference and seamless, cross-device intelligence. The piece frames this shift as a fundamental change in the relationship between humans and technology, moving from interacting with tools to collaborating with intelligent systems. It concludes that the Snapdragon Summit signals Qualcomm's active role in shaping this future, though the journey is just beginning.
Read source