How AutoHome Built the Cangjie Large Model: From Training Architecture to Real-World AI Applications

Introduction

AutoHome, a leading automotive information platform, shares its experience in applying large‑model AI to accelerate digital transformation in the automotive industry. The company leverages extensive private data (vehicle specs, user reviews) to overcome the knowledge gaps of generic models and to provide precise, trustworthy answers.

Advantages and Challenges

Data Barrier & Expertise : Proprietary data gives a competitive edge for complex automotive queries.

User Behavior Insight : Rich user interaction data enables personalized services.

High Professional Requirements : Automotive knowledge is highly specialized, making generic models insufficient.

Training Difficulty : Incorporating domain knowledge into large models demands massive compute resources.

Training Architecture

The training framework is engineered for "no‑code" operation and supports distributed training across multiple GPUs. Three parallelism strategies are employed:

Data Parallelism : Same model on multiple devices processes different data shards.

Pipeline Parallelism : Model layers are split across devices, each handling a portion of the forward pass.

Tensor Parallelism : Individual tensors (e.g., weight matrices) are partitioned across devices.

Implementation builds on industry‑standard tools such as DistributedDataParallel (DDP), DeepSpeed (Zero‑2/Zero‑3), and Megatron‑LLM 3D parallelism. AutoHome also integrates open‑source frameworks tailored to its needs:

MS‑Swift : Multi‑modal training framework.

LLaMA‑Factory : Language‑model‑focused training pipeline.

VERL : Reinforcement‑learning framework.

Business Application 1 – Video Script Generation

The goal is to generate high‑quality, non‑generic video scripts for automotive content. The workflow consists of:

Video script analysis – extract themes and structures from top‑rated videos.

Model fine‑tuning – train on high‑engagement scripts.

Like‑rate prediction – a dedicated model estimates the popularity of generated scripts.

Reinforcement learning – use the like‑rate estimator as a reward to further optimize the generator.

Deployment – feed the final scripts into downstream video production pipelines.

Data construction pairs high‑engagement scripts (answers) with queries generated from the knowledge base, creating a supervised dataset for fine‑tuning. The like‑rate estimator evolved from a simple text classifier to a multimodal model incorporating image and video features, improving correlation by ~10% after adding cover‑image and video‑frame features.

Business Application 2 – Agent Capability

AutoHome built an Agent that can autonomously plan, call tools, and provide answers. The Agent follows a five‑step loop:

Perception : Receive user input.

Planning : Determine which internal tools to invoke.

Decision : Choose the next action based on tool results.

Execution : Call APIs, databases, or other services.

Feedback : Analyze outcomes and decide whether to continue or respond.

Tools are exposed via the MCP (Multi‑Tool Call Protocol) , a standardized function‑calling interface. Example workflow shows the Agent retrieving a vehicle’s transmission type by first identifying the model ID, then querying the configuration service, and finally composing the answer.

Non‑Parametric Optimizations

Prompt Engineering : Role prompts, few‑shot examples, and constraint prompts guide the model.

Context Management : Summarize tool responses to keep token usage low while preserving essential information.

Tool Guidance : Introduce reflection tools to verify critical parameters (e.g., ID validity) and combine tool calls for richer interactions.

Parametric Optimizations (Fine‑Tuning)

Ignore loss from raw tool responses; focus on the model’s reasoning.

Reduce redundant thinking steps during training.

Weight loss higher for tool‑call tokens to improve tool usage.

Mix domain‑specific data with generic data to enhance generalization.

Inference Optimization

Four primary goals guide the serving stack: low latency, high throughput, cost‑effectiveness, and broad compatibility. AutoHome evaluated open‑source inference engines and built a customized solution:

vLLM : Best compatibility across model families.

LMDeploy : Highest raw speed.

SGLang : Scalable for extremely large models.

Key techniques include INT8/INT4 quantization and speculative decoding, which boost throughput by 40%–70% with less than 2% accuracy loss. The optimizations free up GPU capacity for additional training or services.

Summary and Outlook

The Cangjie large model establishes a strong foundation for automotive‑specific AI, while the Agent with MCP closes the perception‑planning‑execution‑feedback loop, delivering more accurate and trustworthy answers. Future work focuses on personalized Q&A by leveraging user behavior profiles to tailor recommendations, and extending the Agent with reinforcement‑learning‑based decision making for end‑to‑end car‑selection assistance.