The parameter scale of on-device smartphone models is undergoing a rapid expansion.
On September 22, Qualcomm partnered with StepFun, Infinite Fire, and Longsys to fully deploy a 30B MoE model on the sixth-generation Snapdragon 8 Ultra mobile platform, completing validation on reference design devices with model inference requiring no cloud connectivity.
This is not an isolated case.
On September 16, vivo announced pre-research into a BlueHeart 30B MoE on-device model. On October 13, ModelBest released the 35B-A3B open-source model MiniCPM-V 4.7 and published actual device running results on the iPhone 17 Pro Max.
In just one month, 30B-class on-device models have appeared one after another. Models with tens of billions of parameters that were once blocked by smartphone RAM and computing power limitations are beginning to show feasibility for deployment.
This shift is occurring precisely against the backdrop of continuously rising smartphone storage costs, and stands in contrast to the direction two years ago when smartphone manufacturers worked hard to slim down on-device models.
In 2024, vivo's key offering was still a 3B model occupying only 1.4GB of RAM. Now, the company has begun exploring a 30B model with ten times the total parameter scale.
The important change behind this is the gradual maturation of technologies such as MoE (Mixture of Experts). By reducing the parameters actually involved in computation during each inference, smartphones can begin to run larger-scale models with relatively limited computing power.
This has driven the large model scaling race, previously conducted mainly in the cloud, to extend toward the smartphone side. Smartphone manufacturers are also beginning to attempt to capture the scale benefits brought by Scaling Laws.
But even as model capacity expands, the RAM, power consumption, and cost budgets that smartphones can allocate to AI have not grown in sync.
This also places the smartphone industry chain in a technology bet full of uncertainty.
On one hand, whether larger models can truly bring about a leap in smartphone AI capabilities and form new selling points that consumers can perceive, or even be willing to pay for, remains unknown.
On the other hand, AI model technology roadmaps may change within just a few months, while smartphone chips typically take about 24 months from planning to deployment. Whether hardware investments made today for large-parameter models can adapt to new technology roadmaps two years from now is equally full of variables.
The Scale Dividend of Smartphone AI
Smartphone manufacturers have not always pursued large-parameter on-device models.
In 2023, vivo announced it had successfully run 1B, 7B, and even 13B on-device large models. Just one year later, the company shifted its focus to the smaller 3B model.
In 2024, vivo launched BlueHeart 3B, with memory usage of only 1.4GB. According to vivo's statements at the time, this model could match the performance of 7B to 9B models on certain tasks, with further reduced power consumption compared to the earlier BlueHeart 7B.
This was also an important direction for the smartphone industry at the time. In 2024, the Honor Magic6 series was equipped with a 7-billion-parameter on-device large model, while Apple's Apple Intelligence adopted an approximately 3-billion-parameter on-device foundation model.
By 2025, vivo continued to release 3B on-device multimodal reasoning models; Apple's upgraded on-device foundation model that year also maintained a parameter scale of approximately 3B.
Zhou Wei, Vice President of vivo and Dean of vivo AI Global Research Institute, once believed that on-device large models should not blindly pursue parameter scale, and that the key lies in meeting user needs. He even judged that lightweight deployment would become an industry trend.
This judgment had practical considerations. Smartphone RAM also needs to be reserved for the operating system, camera, games, and other applications. How to improve model capabilities within limited memory and power budgets was an important goal for smartphone manufacturers at the time.
But this year, Zhou Wei began to revise his earlier judgment.
"Last year, we felt that running 3B and 7B models on smartphones was very demanding on memory," Zhou Wei candidly told All-Weather Technology. At the beginning of this year, new technology changes emerged in the industry.
The key driver of this change is that technologies such as MoE have partially decoupled the total parameter scale of a model from the computation required for a single inference. Traditional dense models typically need to invoke all model parameters during each inference, while MoE divides part of the network into multiple experts, selecting only a few experts to participate in computation each time.
Taking the 30B-A3B model as an example, the total parameter count reaches 30 billion, but when processing each token, the parameters actually involved in computation number approximately 3 billion.
This means that a model can increase its capacity by expanding total parameter scale, without actual computation growing proportionally.
In fact, MoE has long been an important technology roadmap for scaling cloud-based large models. Over the past few years, companies including DeepSeek and Alibaba have successively released large models adopting MoE architecture, demonstrating the potential of sparse models in expanding capacity while controlling computational costs.
Now, this approach is beginning to extend to smartphones.
Yao Yuan, Co-Chief Scientist of ModelBest and Assistant Professor at Tsinghua University, explained to All-Weather Technology that the important value of MoE lies in achieving decoupling of model capacity from computational cost.
He compared it to a library: users can enjoy the knowledge reserves of an entire library, but each time they answer a question, they only need to access content from a few key bookshelves.
With the development of MoE and related deployment technologies, manufacturers can also store some expert weights in flash memory, loading them into RAM only when needed, reducing the model's resident memory requirements.
Zhou Wei believes that with flagship smartphones now commonly equipped with 12GB of RAM, allocating approximately 3GB to run MoE models without affecting normal use is becoming feasible.
"We revised our judgment on this technology from last year, as well as the industry's overall technology judgment, this year, and then quickly developed our own MoE model," Zhou Wei told All-Weather Technology.
This has also led smartphone manufacturers to reconsider the scale benefits brought by Scaling Laws.
In the past, the large model industry continuously improved model capabilities by increasing model parameters, training data, and computational investment. Constrained by hardware conditions, smartphone manufacturers long focused on technologies such as model compression, quantization, and knowledge distillation to reduce deployment costs.
Now, MoE has opened up space for smartphones to expand model capacity, giving manufacturers the opportunity to gain capability improvements from larger models at limited computational cost.
However, a 30B MoE and a 30B dense model cannot be simply compared by total parameters. Although the former has greater model capacity, the parameter scale actually activated each time remains only a few billion, and its capability improvement is not equivalent to running a 30B dense model.
Greater model capacity may also not bring proportional performance improvements across all tasks.
How much of the scale benefits from Scaling Laws can actually be realized on the smartphone side still needs to be verified under specific task, memory, and power constraints.
What More Can Smartphones Do?
Compared to the commonly seen 3B and 7B on-device models of the past, what capability improvements a 30B MoE can actually bring is the question truly worth paying attention to in this round of parameter expansion.
Yao Yuan believes that MoE's advantages are mainly reflected in tasks that rely more on knowledge breadth and representation capacity, such as multimodal perception, OCR and document understanding, STEM reasoning, and long video comprehension.
Taking ModelBest's newly released MiniCPM-V 4.7 35B-A3B as an example, according to test data published by ModelBest, it achieved scores of 91.2, 91.72, and 86.45 on OCRBench, MMBench, and Mathverse_MINI respectively, leading among models of the same level and even surpassing larger models.
This also corresponds to the application changes occurring in smartphone AI.
In the past, on-device smartphone models mainly handled relatively independent tasks such as text summarization, simple Q&A, and photo recognition. But as smartphone AI evolves toward personal agents, the content models need to understand is becoming increasingly complex, including photos, videos, emails, calendars, and users' long-accumulated personal information.
Models also need to formulate plans and invoke smartphone system functions based on understanding this information.
At the Snapdragon Summit in September this year, Qualcomm demonstrated such a scenario: on a reference design device equipped with the sixth-generation Snapdragon 8 Ultra mobile platform, the StepEdge-Omni 30B-MoE model could read an invitation email containing dates, flight numbers, and event locations, understand the information within, and then combine it with the calendar to complete a series of tasks including itinerary planning, hotel and flight recommendations.
The entire model inference process was completed locally.
According to data published by Qualcomm, after joint optimization, this solution achieved prefill throughput exceeding 330 tokens per second and decode throughput exceeding 28 tokens per second, demonstrating the interaction potential of 30B-class models on smartphone-level hardware.
For smartphone manufacturers, keeping more powerful models on-device has another layer of appeal.
Users' photos, emails, chat records, and calendars — large amounts of personal information — are already stored on smartphones. If models can continuously understand this data locally, there is an opportunity to reduce reliance on cloud transmission and inference, and lower response latency for some tasks.
But the question is, to what extent do these capabilities actually need to depend on a 30B MoE model?
Past 3B and 7B models have already been able to handle some tasks such as email understanding, calendar planning, and photo retrieval. Small models, through coordination of multiple specialized models and integration with external tools, also have the opportunity to complete some complex work.
Yao Yuan also acknowledged that MoE's advantages have a scope of applicability. On tasks such as pure text code, complex instruction following, and GUI localization, MoE models with 3B activated parameters and 7B dense models still trade wins and losses.
"MoE does not equal 'getting 30B for free,'" he said, noting that the actual activated parameter count still limits the capability ceiling for some tasks.
Whether agents can truly complete cross-application operations also depends on whether the smartphone operating system opens corresponding permissions and the reliability of model tool invocation.
Currently, the test environments, quantization schemes, and task types of 30B models demonstrated by various manufacturers are not entirely consistent, and there is a lack of sufficiently unified real-world scenario comparisons. In particular, under similar computing, memory, and power budgets, how much a 30B MoE can improve the success rate of complex tasks compared to 3B and 7B models still lacks sufficient verification.
For smartphone manufacturers, being able to run a 30B MoE model on a smartphone is only the first step.
Whether larger models can accomplish tasks that were difficult for smaller models in the past, or significantly improve AI experiences that users can perceive, and further translate into reasons for consumers to purchase next-generation smartphones, is what determines the commercial value of this round of parameter expansion.
The Hardware Ledger
Even though MoE reduces the computation required for a single inference, the complete model weights still need to be stored and read.
Taking a 30B MoE model as an example, even with 4-bit quantization, the weights corresponding to 30 billion parameters alone theoretically require approximately 15GB of storage space, with actual usage also affected by quantization methods, model architecture, and other operational data.
For smartphones, these model weights can be stored in flash memory, but it is difficult to load them all into RAM for extended periods.
To address this, manufacturers have begun exploring storing most expert weights in flash memory and loading only the portions needed for current inference into RAM.
This approach reduces the model's RAM capacity requirements, but brings a new problem: if the required experts have not yet been loaded, the computing unit must wait for data reading.
"Fewer activated parameters only saves computation; the 35B weights themselves still need to be stored and read," Yao Yuan explained to All-Weather Technology, noting that the important bottleneck for current on-device MoE lies in memory capacity and storage bandwidth.
ModelBest's release of MiniCPM-V 4.7 this time focused on optimizing precisely this aspect. Its LES (Lookahead Expert Scheduling) technology can predict which experts may need to be invoked in the next layer while the current layer is computing, and read the corresponding weights in advance, allowing some data loading to proceed in parallel with model computation.
According to test data published by ModelBest, on an iPad Pro M4 equipped with 16GB of RAM, LES enabled the model to achieve an average decoding speed 5 times faster than before optimization.
ModelBest also introduced DSpark speculative decoding technology, which further reduces the repetitive overhead of token-by-token generation by verifying multiple candidate tokens at once. In tests on the iPhone 18 Pro Max, MiniCPM-V 4.7 35B-A3B used only about 3GB of RAM and achieved a decoding speed of over 20 tokens per second.
These advances indicate that the efficiency of running large-parameter models on smartphones increasingly depends on the collaborative capabilities of the entire storage and computing system.
Chip manufacturers are also adjusting their architectures for this.
In September this year, Qualcomm announced a redesigned Hexagon NPU. One important change was expanding on-chip shared memory of the NPU by 50%, allowing more model data and intermediate computation results to remain within the chip, reducing external data transfer.
Qualcomm also introduced an Element Accelerator dedicated acceleration unit for Transformer models, and supports multiple computation precisions from INT2 to FP16.
In the 30B MoE model deployment, Qualcomm further collaborated with companies including Infinite Fire and Longsys on coordinated optimization around inference engines, heterogeneous computing, and storage data reading.
According to data published by Qualcomm, the optimized model's RAM requirements were reduced by more than 50% compared to conventional solutions.
Yao Yuan believes that next-generation smartphone chips need further optimization around MoE's sparse activation, expert loading, and low-bit computing, including expanding shared memory available for AI computing, improving chip cache and data transmission pathways, and enhancing the collaborative efficiency between model weight reading and computation.
"What wins on the device side is 'channels and coordination' — that is, bandwidth and collaboration," Yao Yuan said.
But these technology investments are occurring precisely during a period of continuously rising smartphone storage costs. Cost pressure has even affected smartphone manufacturers' configuration strategies. TrendForce noted that in 2026, high-end smartphone RAM configurations will begin to concentrate more on 12GB, with 16GB configurations being compressed.
On one hand, smartphone manufacturers are working to control memory costs; on the other hand, on-device models are expanding toward 30B and even larger parameter scales. How to balance model capability with hardware investment has become an issue manufacturers must consider.
If running large-parameter MoE models requires higher storage costs, more complex data scheduling, and additional power expenses, can smartphone manufacturers obtain product premiums through better AI experiences?
If local models can handle more inference tasks, can they also reduce the cloud invocation costs that manufacturers continuously pay?
Currently, neither of these calculations has a clear answer.
Moreover, frequent reading of model weights will also increase the smartphone's energy consumption pressure. Zhou Wei told All-Weather Technology and other media that compared to keeping expert weights resident in memory, continuously moving model weights may bring greater power consumption pressure.
He revealed that when vivo's earlier 3B model was first launched, its operating current reached over 800 milliamps, and after a year of optimization, it was reduced to 500 to 600 milliamps.
For 30B MoE models, Zhou Wei considers below 500 milliamps to be a relatively ideal operating target; by the end of 2028, if the operating current can be controlled at around 600 milliamps, it is expected to have practical usage conditions.
Smartphone internal space is compact, and heat dissipation capability is limited. Especially as future personal agents need to continuously process personal information and execute tasks in the background, the power consumption and heat generation from long-term operation will directly affect the actual experience of related functions.
Zhou Wei estimates that even by the end of 2028, flagship smartphones equipped with 12GB to 16GB of RAM, with the memory budget truly allocated to on-device models controlled within 4GB, will still be a relatively reasonable state.
This means that model manufacturers need to continue optimizing inference efficiency, while chip and storage manufacturers need to improve the collaborative capability between data scheduling and computation.
At the same time, future changes in model technology roadmaps bring another layer of uncertainty to hardware investment.
Today, MoE provides a new path for smartphones to run models with tens of billions of parameters, and chip and storage manufacturers have begun optimizing for the operating characteristics of sparse models.
But whether future smartphone AI will rely on a single MoE model with greater capacity, or complete tasks through coordination of multiple small models, still has different technology options.
When introducing the new-generation Hexagon NPU architecture, Qualcomm also proposed a development direction: future agents can dynamically select different specialized models based on tasks and context. Small models handle more local tasks, invoking larger models only when necessary.
Different model operation approaches may also have different requirements for chip resources.
If models continue to develop along the large-parameter MoE path, the importance of expert loading efficiency, storage bandwidth, and cache scheduling will further increase.
If more efficient small model architectures emerge in the future, or if solutions coordinating multiple specialized models perform better in practical applications, smartphone requirements for model capacity, data movement, and computing resources may also change accordingly.
Of course, hardware optimizations around MoE also have certain general value. Larger shared memory, more efficient data transmission pathways, and low-bit computing capabilities can also serve other model architectures and multimodal tasks.
Moreover, chip manufacturers must make these choices in advance during a period when model technology is still rapidly changing.
Zhou Wei candidly stated that a smartphone chip typically takes about 24 months from planning to deployment.
This means that hardware planning made today around 30B MoE may not be fully realized until one to two chip generations later.
But two years is enough time for multiple rounds of technology changes in the AI industry.
In 2024, smartphone manufacturers were still working hard to make models smaller; in 2026, 30B has become a new exploration direction. By 2028, whether the model architectures receiving attention today will still be suitable for next-generation smartphones remains uncertain.
Models can change direction quickly, but chips must place their bets two years in advance.
As the scale benefits of Scaling Laws begin attempting to extend to the smartphone side, large-parameter models have already passed the threshold of preliminary operational validation.
Next, smartphone manufacturers need to prove that larger models can bring capability improvements worth consumers paying for, while chip and storage companies need to ensure that today's hardware investments can adapt to continuously changing technology roadmaps in the future.
This technology bet still has no answer.
Comments