OPPO’s On-Device Multimodal LLM Engineering Practice at AICon Shenzhen

OPPO’s Dr. Xiyu Yu is confirmed to speak at the AICon Shenzhen 2026 conference on the topic of on-device multimodal large language model (LLM) engineering practice. His talk will address core challenges such as prefill/decoding latency in long-context scenarios, complex engineering pipelines from training compression to deployment, and difficulties in model iteration and personalization. OPPO’s full-stack approach includes a modular QAT training framework integrating sparse training, eviction-aware QAT, and decoding acceleration. An automated QALFT pipeline covers data preparation, PTQ/QAT, QALFT fine-tuning, accuracy validation, and performance profiling across MTK and Qualcomm platforms. Zero-order optimization using sparse perturbation enables on-device training for privacy-preserving personalization. The practice has been applied to over 10 vertical business scenarios, supporting rapid iteration of on-device agent capabilities. Future directions include efficient attention and KV cache management for longer contexts, lower bit-width quantization, and unified optimization of multimodal understanding and agent planning. The conference will take place on August 21-22, 2026, featuring 50+ experts from top tech companies.