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

This article introduces Vivo's BlueLM large model, presented by Vivo Algorithm Director Fu Fan, outlining the model's evolution from expert systems to modern large language models and highlighting OpenAI's breakthroughs that paved the way for general AI.

The discussion then moves to the training and challenges of BlueLM, emphasizing massive data handling, efficient algorithm development, and safety‑controlled practices, including data collection, cleaning, deduplication, sampling, and multi‑stage algorithm design (pre‑training, SFT, reinforcement learning, and prompt engineering).

BlueLM employs a model matrix ranging from 1B to 175B parameters, balancing task performance, inference cost, and hardware constraints; smaller models run on‑device while larger ones serve cloud services. The architecture follows a "large‑and‑complete" principle, supporting diverse modalities and parameter scales.

Key technical highlights include mixed‑precision training, gradient scaling, loss prediction, target‑only loss for fine‑tuning, clustering‑based data balancing, and safety‑oriented reward models. Edge‑side optimization combines model compression, quantization, and platform‑specific frameworks to achieve real‑time performance on smartphones (e.g., 60 tokens/s for 1B model, 20 tokens/s readable speed).

Productization is demonstrated through BlueLM Mini‑V, an on‑device conversational assistant deployed on Vivo X100, iQOO12, and S18 series, and BlueQianXun, an Android/iOS app offering voice interaction, Q&A, and image generation. Interactive features such as floating widgets and image editing are also described.

The article concludes with a Q&A segment where Fu Fan answers ten technical questions about model‑API integration, full‑parameter vs. LoRA fine‑tuning, data‑type ratios, sampling strategies, safety mechanisms, and the rationale for training from scratch rather than using existing open‑source models.

Overall, the presentation showcases Vivo's end‑to‑end workflow from large‑model research to real‑world AI product deployment, emphasizing scalability, safety, and user‑centric performance.