Skip to main content

Large Model Lightweighting: Mini Versions New Trend

Introduction: Large Models Are Getting "Smaller"​

WAIC 2026 has just concluded, and large models from manufacturers seem to have changed.

At last year's conference, companies competed on model parameters and benchmark scores. This year, focus is no longer on "how big the model is" but "what can the model do and can it make money?" Manufacturers are all pursuing lightweight deployment and developing mini versions.

From Mobile Phones​

Samsung Galaxy S24 series featured "Circle to Search" AI function. Behind this are Qualcomm Snapdragon and MediaTek chips promoting 10-billion-parameter large models that can run locally.

Local running of large models has special meaning. The photos you take, your account screenshots, chat records - this data should never be uploaded to any cloud. Smartphones have become our most personal digital organs.

From Automobiles​

When end-to-end AI large models control autonomous driving real-time decisions, every millisecond matters. A car must be a wheeled mobile intelligent agent that can think independently. Its "brain" must be on the car itself.

From Factories​

Industrial intelligence is sinking from central server rooms to every robotic arm, every drill bit, every sensor. Lightweight visual inspection models run directly on edge computing modules.

Technical Paths: Four-Step "Slimming"​

Quantization​

Representing numbers in shorter formats: from 32-bit floating-point (FP32) to 16-bit (FP16) or even 4-bit (INT4), exchanging information for space.

Distillation​

Training a small "student model" to imitate how the huge "teacher model" thinks and outputs answers, compressing vast knowledge systems into digestible essence.

Pruning​

Removing "redundant" parameters that contribute little to results, like laying off underperforming employees, keeping only the core working backbone.

Structural Optimization​

Not using "general large models" but targeting specific scenarios, creating small and specialized "expert models" with knowledge graphs.

Commercial Value of Mini Versions​

  • Cost: Zero local cost vs. cloud calls
  • Latency: Millisecond local response vs. hundreds of milliseconds cloud
  • Privacy: Data stays local, solving compliance issues

Industry Landscape​

Chip Manufacturers​

NVIDIA, Qualcomm, MediaTek actively promote this movement. If large models run on end devices, they sell more chips.

Mobile Phone Manufacturers​

Apple, Xiaomi, vivo, OPPO treat lightweight large models as new differentiation battlefield. Large models on phones bring revolutionary interactive experiences.

Model Manufacturers​

OpenAI, Anthropic, Google, Baidu, Zhipu AI are at the core. Competition shifted from parameter size to who has smaller, faster, cheaper models running smoothly on end side.

Cloud Manufacturers​

Alibaba Cloud, Tencent Cloud, AWS provide "cloud-edge-end" integrated hybrid deployment solutions, transforming from computing power suppliers to intelligent computing schedulers.

Future Outlook​

Large model lightweighting is not technological regression but technological maturity.

Prediction 1: Model Capability Tiering Becomes Standard​

Future AI services will have model families: giant cloud super brains, medium edge models, countless tiny end-side micro models.

Prediction 2: End-side Intelligence Spawns New Scenarios​

When smart AI can run offline everywhere, new application scenarios will emerge: elderly health monitoring, agricultural drones, factory workers' AI assistants.

Prediction 3: Computing Power Equality Era Arrives​

Lightweighting breaks the threshold. Ordinary developers and SMEs can run capable large models without astronomical computing clusters.

Of course, challenges remain. But the wheel of history rolls forward. Large models once shocked the world with their huge bodies. Now, in a lighter, more flexible way closer to each of our lives, they are truly and profoundly changing every corner of this world.

Source: 36Kr · July 22, 2026