Qualcomm Launches Two AI Smartphone Chips: Flagship Model Can Run 30B-Parameter MoE Model Locally

Qualcomm has unveiled two new smartphone chips with on-device AI as the central selling point. The most notable claim: the new flagship chip can run a 30 billion parameter (30B) Mixture-of-Experts (MoE) model locally. If this holds up in practice, it would represent a significant leap forward in mobile AI inference capabilities.

What Does Running a 30B MoE Model On-Device Actually Mean?

Historically, running large models with billions of parameters required the massive compute power of cloud data centers. Qualcomm's claim that its flagship chip can run a 30B-scale MoE model locally is significant because it moves what was once server-side AI workload directly onto the device in a user's hand.

The MoE (Mixture-of-Experts) architecture is itself a key technique for improving model efficiency. Rather than activating all 30 billion parameters on every inference pass, it uses a routing mechanism to call on only a subset of "expert" networks. This means that while the total parameter count is 30B, the number of parameters actually activated during any single inference is much smaller — enabling higher efficiency within the constrained compute and memory environment of a mobile device. This is precisely why MoE models have attracted growing interest for on-device deployment in recent years.

Looking at the historical evolution of model scale, mainstream on-device AI models have long hovered below 7B parameters. Around 2023, smartphone chips could generally only run models of 1B to 3B parameters smoothly; by 2024, some flagship chips pushed that ceiling to 7B–13B. The 30B parameter range was previously the exclusive domain of cloud services — GPT-3.5 sits at roughly 200B parameters, and Meta's Llama 2 13B was already considered a fairly large limit for on-device deployment. If Qualcomm's claim of locally running a 30B MoE model is validated on real hardware, it would represent an order-of-magnitude leap in mobile model scale within just two years. This leap relies not only on raw chip performance gains, but also on the maturation of low-bit quantization techniques (such as INT4/INT8 compression) — quantization can compress the storage footprint of a 30B model to as little as one-quarter of its original size, making it feasible to fit within a phone's limited memory.

Why Qualcomm Is Betting on On-Device AI

Running AI inference locally on the device brings immediate benefits in privacy, low latency, and offline availability. Data doesn't need to be uploaded to the cloud, keeping sensitive user information on the device itself. Local inference also eliminates network round-trip latency, enabling smoother real-time interactive experiences.

From a competitive standpoint, on-device AI is becoming the new battleground for smartphone chip makers. As generative AI applications become more widespread, whoever can run larger models more efficiently on a phone will hold a significant advantage in the next generation of intelligent device experiences. Qualcomm's decision to center the marketing of both new chips around AI is a direct response to this trend.

How the Two New Chips Are Positioned

Launching two chips simultaneously typically signals targeting of different market tiers. The flagship model serves as a showcase of the technology's upper ceiling — locally running a 30B MoE model is the embodiment of that peak capability — while the second chip is more likely aimed at the upper-mid-range market, striking a balance between AI performance and cost.

This tiered strategy helps Qualcomm extend on-device AI capabilities across a broader range of device models, not just top-end flagships. For device manufacturers and developers, it also means AI applications can be designed and optimized for different performance tiers.

The Real-World Challenges of On-Device Large Models

As exciting as running a 30B-parameter model locally sounds, significant practical challenges remain. A mobile device's memory capacity, thermal management, and battery life all impose constraints on sustained large model operation. While the MoE architecture reduces the number of parameters activated per inference, storing the complete model still requires considerable on-device space.

Moreover, the actual user experience of running a large model locally — including inference speed, response quality, and battery drain — still needs to be validated on real hardware. There is often a meaningful gap between a vendor's claim of "can run" and the smooth, responsive experience users expect in daily use. These are the key factors to scrutinize when evaluating the AI capabilities of these two chips.

Memory bandwidth is a frequently overlooked but critical bottleneck in on-device large model inference. The performance ceiling for large model inference is often determined not by a chip's peak compute (FLOPS), but by how quickly it can move model weights from memory into the compute units — a metric known as memory bandwidth. Mobile SoCs typically deliver memory bandwidth in the 50–100 GB/s range, far below the hundreds of GB/s offered by data center GPUs. Even after quantization compresses a 30B model down to roughly 15–20 GB, generating each token still requires reading all or most of the activated weights, placing sustained pressure on bandwidth and directly impacting generation speed. The MoE architecture can theoretically ease this pressure — activating only a subset of experts per inference means fewer weights need to be moved — but the routing mechanism itself introduces additional scheduling overhead, and real-world gains need to be confirmed through testing on specific hardware.

Conclusion

Qualcomm's two new chips send a clear signal: on-device AI has moved from concept to core product strategy for chip makers. The flagship model's ability to locally run a 30B MoE model sets a new benchmark for mobile AI inference. As more devices powered by these chips come to market, we'll have the opportunity to see how on-device large models perform in real-world scenarios — and whether they can genuinely transform the AI experience on smartphones.