PRIME A new way to shrink AI models without losing smarts

The problem with giant AI models

Modern AI chatbots and writers are built from Mixture-of-Experts (MoE) models—a design that splits a huge network into many smaller “expert” modules. When a question arrives, the model picks the experts that are most relevant, which makes inference faster and cheaper than using one monolithic block. However, even though only a fraction of parameters run at any moment, the total number of experts can still be enormous, demanding massive GPU memory and slowing down deployment.

Existing compression techniques try to solve this by either pruning away unused experts or merging similar ones, but they treat these steps separately. That disjoint approach often throws away useful information or mixes incompatible expertise, leading to a drop in performance that is hard to recover.

How PRIME works

PRIME (Frequency-Aware Unified Pruning and Merging for MoE Compression) changes the game by doing pruning and merging in a single, coordinated pass. Instead of making greedy, one-by-one decisions, PRIME frames the whole process as an integer linear programming problem—a mathematical optimization that finds the best overall assignment of each expert to either a pruned or a merged group.

The key insight is frequency awareness. The method tracks how often each expert is called during normal use. Experts that see high traffic become the stable “centroids” that others merge into, while rarely used experts are more likely to be cut. This prioritizes routing stability, because the model continues to rely on the most frequently needed pathways.

Results on real models

The researchers tested PRIME on two large Alibaba models—Qwen3-30B-A3B and Qwen3.5-122B-A10B. After compression, the models retained strong performance across a range of tasks: general knowledge, coding, mathematics, and long-context reasoning. When the compressed models were served using the open-source vLLM engine, benchmarks showed significant gains in memory efficiency and throughput, proving that the method can turn bloated MoE giants into leaner, production-ready systems.

Why it matters

As AI models keep growing, the pressure to make them smaller and faster only intensifies. PRIME offers a principled way to shrink MoE architectures without the costly trial-and-error of ad-hoc pruning or merging. By unifying both steps and weighing expert usage, it helps bridge the gap between research-grade scale and real-world deployment, paving the way for cheaper, quicker AI services that don’t sacrifice intelligence.

Key facts

PRIME jointly prunes and merges MoE experts in a single optimization step.
It uses frequency‑aware integer linear programming to keep high‑traffic experts as stable centroids.
Tests on Alibaba's Qwen3‑30B‑A3B and Qwen3.5‑122B‑A10B show strong accuracy retention.
Compressed models served with vLLM demonstrate clear memory and throughput improvements.
The method tackles the memory bottleneck of large MoE models without disjoint, greedy compression.