From BigGo Finance · BigGo Finance · · 2m
30B-Class Models Squeezed Into Phones
On-device AI arms race pushes smartphone memory to new limits.
In brief
On-device AI arms race pushes smartphone memory to new limits. Smartphone makers are racing to deploy 30B-class AI models on devices, but face memory and power constraints as storage costs rise. Originally reported by BigGo Finance.
30B-Class Models Deployed on Smartphones
Over the past month, 30B-class mixture-of-experts (MoE) models have successively completed deployment validation on smartphone platforms.
Qualcomm and Partners Deploy 30B MoE Model
Qualcomm, together with StepFun and other partners, ran a 30B MoE model on the Snapdragon 8 Elite Gen 6 platform.
Vivo Begins BlueLM 30B Research
Vivo has begun preliminary research on its BlueLM 30B on-device model.
ModelBest Releases MiniCPM-V 4.7
ModelBest released the open-source 35B-A3B model MiniCPM-V 4.7 on October 13.
MoE Architecture Decouples Parameters
MoE partially decouples a model's total parameter count from the compute required per inference.
MoE Activates Fewer Parameters Per Inference
A 30B-A3B model uses only about 3 billion parameters per inference instead of all 30 billion.
Yao Yuan Compares MoE to Library
Yao Yuan likens MoE to a library where users draw on knowledge reserves but only consult a few bookshelves per question.
MoE Models Show Strength in Specific Tasks
MiniCPM-V 4.7 scored 91.2 on OCRBench, 91.72 on MMBench, and 86.45 on Mathverse_MINI benchmarks.
Qualcomm Demonstrates On-Device Capabilities
StepEdge-Omni 30B-MoE can read emails, understand content, and complete itinerary planning locally.
Memory Constraints for 30B Models
A 30B MoE model requires about 15GB of storage space even with 4-bit quantization.
ModelBest Uses LES Technology
ModelBest introduced LES (look-ahead expert scheduling) to improve decode speed and memory usage.
Qualcomm Optimizes Hexagon NPU
Qualcomm's redesigned Hexagon NPU expands on-chip shared memory by 50% and adds an Element Accelerator.
Rising Storage Costs Challenge Vendors
TrendForce notes that flagship phone memory configurations will trend toward 12GB in 2026, with 16GB configurations squeezed.
Power Consumption Targets
Vivo aims for under 500 milliamps for 30B MoE models by 2028, down from 800 milliamps for 3B models.
24-Month Chip Development Lag
Smartphone chips take about 24 months from planning to launch, creating uncertainty for 30B MoE deployments.
The essence
Smartphone makers are racing to deploy 30B-class AI models on devices, but face memory and power constraints as storage costs rise.
