How JD’s Omniforce Boosts Large Model Efficiency with Cloud‑Edge Collaboration

Research Overview

On May 15, the JD Exploration Institute published the paper “Omniforce: A Human‑Centred, Empowering Large‑Model, Cloud‑Edge Collaborative AutoML System” in the Nature‑partner journal npj Artificial Intelligence . The study proposes a system and methodology for training and updating large models in open‑environment scenarios while coordinating with smaller models for deployment.

Key Innovations

Model Distillation: Dynamic hierarchical distillation, especially during pre‑training, adjusts only 0.5% of parameters to achieve efficient training in low‑resource settings, reducing deployment cost of large models.

Data Governance: A cross‑domain dynamic sampling algorithm automatically mixes data from different domains, incorporating privacy‑preserving and active‑learning techniques to enhance model generalisation.

Training Optimisation: Bayesian optimisation (BO) drives hyper‑parameter tuning and architecture search, excelling in discrete spaces and raising resource utilisation by 40% in MPMD scenarios.

Cloud‑Edge Collaboration: The cloud conducts model search and training, while edge devices handle deployment and inference; a two‑stage compression strategy adapts to resource constraints and improves cloud‑edge efficiency.

Impact and Platform Integration

The proposed techniques raise average inference efficiency by 30% and cut training cost by 70%, offering a reusable technical paradigm for scaling large‑model deployment. These advances underpin the JoyBuild large‑model development platform, which supports JD’s own models as well as Llama, DeepSeek and others.

JoyBuild provides over 20 open‑source models, 100+ algorithms and toolchains, and a rich dataset library, enabling enterprises to transform generic models into specialised solutions with a dramatically reduced team size—from ten‑plus scientists to just one or two algorithm engineers.

Extensive experiments confirm the effectiveness: inference speed improves by 30% on average, and training expenses drop by 70%. For example, after distillation, JD’s large model Livebench gains 14 points over comparable models.

Broader Industry Relevance

JoyBuild’s technology is already applied across JD’s retail, logistics, health, and finance domains, supporting use cases such as supply‑chain optimisation, intelligent customer service, and marketing content generation. The approach is presented as an open, reproducible pathway rather than a black‑box solution, offering valuable reference for both academia and industry.