How JD’s Omniforce Cuts Large‑Model Training Cost by 70% and Boosts Inference Speed 30%

The research paper "Omniforce: A Human‑Centred, Large‑Model‑Empowering, Cloud‑Edge Collaborative AutoML System" was published in the Nature‑affiliated journal npj Artificial Intelligence, marking the first domestic system to systematically address the efficiency challenges of developing large models in open environments.

Key Performance Gains

The proposed technology achieves an average 30% improvement in inference efficiency and a 70% reduction in training cost, providing a reusable paradigm for large‑model deployment.

Four Core Innovations

Model Distillation: Dynamic hierarchical distillation, especially applied during the pre‑training stage, adjusts only 0.5% of parameters to enable efficient training in low‑resource scenarios, significantly lowering deployment costs.

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 Optimization: Bayesian optimisation (BO) is used for hyper‑parameter tuning and architecture search, excelling in discrete spaces and boosting resource utilisation by 40% in multi‑process‑multiple‑device (MPMD) scenarios.

Cloud‑Edge Collaboration: Model search and training are performed in the cloud, while edge devices handle deployment and inference; a two‑stage compression strategy adapts to resource constraints and improves cloud‑edge efficiency.

Platform Implementation – JoyBuild

The JoyBuild large‑model development platform incorporates these innovations and supports a variety of models, including JD’s own large model, Llama, and DeepSeek. In model‑distillation benchmarks, the JD model’s Livebench score improves by 14 points compared with peers.

JoyBuild offers over 20 open‑source models, more than 100 algorithms and toolchains, and enables end‑to‑end workflows—from data preparation and model training to deployment—within a week, reducing the required team size from ten+ scientists to just one or two algorithm engineers and cutting inference costs by up to 90%.

The platform also embeds JD’s industry knowledge (retail, logistics, health, finance) to accelerate domain‑specific applications such as supply‑chain optimisation, intelligent customer service, and marketing content generation.

Broader Impact

By exposing the underlying methods rather than a black‑box solution, JD’s approach provides a general pathway for improving large‑model training efficiency and applicability, offering valuable reference for both academic research and industrial practice.