Overview of Vivo BlueLM Large Model: Evolution, Training Challenges, and Product Deployment

The talk, led by Vivo's Algorithm Director Fu Fan, introduces the BlueLM (Blue Heart) large model, outlining its development from early expert systems to modern deep learning breakthroughs such as OpenAI's GPT series.

It then details the specific challenges encountered in training BlueLM, including handling petabytes of multimodal data, designing efficient algorithms, ensuring safety and controllability, and achieving high performance on both cloud and mobile devices.

A technical panorama is provided, describing the model matrix (1B, 7B, 70B, 130B, 175B parameters), the end‑to‑end data pipeline (collection, cleaning, deduplication, sampling), and the training stages (pre‑training, SFT, reinforcement learning, prompt engineering) along with optimization techniques like mixed‑precision training and model compression.

The article also showcases product deployments: the BlueLM-powered "Blue Heart Small V" assistant on Vivo smartphones, the "Blue Heart QianXun" cross‑platform app for intelligent Q&A and content creation, and various use‑case examples ranging from work‑related report generation to daily cooking assistance.

Finally, a Q&A segment addresses practical concerns such as model‑parameter selection, data‑sampling strategies, safety review processes, and the rationale for training from scratch rather than adopting existing open‑source models.