Reproduce “AndesVL Technical Report: An Efficient Mobile-side Multimodal Large Language Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhiwei Jin, Xiaohui Song, Nan Wang, Yafei Liu, Chao Li, Xin Li, Ruichen Wang, Zhihao Li and 32 more
The OPPO AI Center developed AndesVL, a family of efficient multimodal large language models designed for mobile-side deployment, achieving first-tier performance across 32 open-source benchmarks. The 4B-Thinking variant surpassed InternVL3.5-4B by 3.2 points, while on-device optimizations yielded up to a 6.7x decoding speedup and 30.9% memory reduction on MediaTek Dimensity 9500 chips.
Key insights
- Integrating long Chain-of-Thought (CoT) data and sophisticated reinforcement learning (MPO/GRPO) in a multi-stage training pipeline is highly effective for enhancing reasoning and instruction-following in compact MLLMs.
- A Native Resolution ViT architecture with 2D-RoPE and meticulous multi-modal data curation across various domains (including mobile-specific UI tasks) is critical for achieving robust performance in mobile environments.
- Advanced optimization techniques like Quantization-Aware LoRA Fine-Tuning (QALFT) and proprietary cache eviction (OKV) are key to enabling flexible, efficient, and high-performance deployment of MLLMs on resource-constrained mobile hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.