China Mobile Unveils Jiutian 3.0 AI Model with Open-Source Tech
China Mobile Debuts Jiutian 3.0 AI Model with Major Performance Gains
At the 2025 World Artificial Intelligence Conference, China Mobile unveiled its Jiutian Foundation Large Model 3.0, marking a significant milestone in AI industrialization. The model builds on its predecessors with enhanced security, controllability, and domestic production capabilities.
Jiutian Zhongqing Language Model: Architecture and Performance Breakthroughs
The Jiutian Zhongqing Language Large Model adopts a MoE architecture, scaling to trillions of parameters. With 15T tokens of pre-training data and a hierarchical post-training framework, it improves complex reasoning by 35%. Key achievements include:
- GPQA-Diamond: Scored 77.67, ranking second globally.
- ArenaHard V1.0: Scored 67.2, topping global rankings.
- BFCL V3: Achieved 68, showcasing robust performance.
The model also addresses hallucinations in professional scenarios, ensuring precise outputs.
Specialized Models for Code and Math
China Mobile introduced specialized models:
- Jiutian Code Large Model: Supports 10+ programming languages, excelling in code generation tasks.
- Jiutian Mathematics Large Model: Outperforms competitors like Qwen2.5Math and DeepSeek Math in SOTA benchmarks.
Jiutian Shanzhi Multimodal Model: Enhanced Visual Capabilities
The Jiutian Shanzhi Multimodal Large Model improves generative controllability and understanding:
- Image Understanding: Scores 82.2 (MMStar), 64.3 (HallusionBench), and 94.9 (OCRBench).
- Video Understanding: Leads in Videomme and MVbench tasks.
The model supports multi-turn precise edits, such as background changes or element additions.
Open-Source Initiatives to Drive Industry Growth
China Mobile open-sourced several technologies to spur innovation:
- Structured Data Large Model (JT-DA-8B): Includes weights and inference code.
- Mathematics Large Model (JT-Math-8B): Available with technical reports.
- Code Large Model (JT-Coder-8B): Supports developer customization.
- TReB Evaluation System: First-of-its-kind structured data benchmarks.
- CCR-Bench Dataset: Simulates industrial scenarios like healthcare and customer service.
These resources are accessible on GitHub, HuggingFace, and other platforms.
Key Points
- Jiutian 3.0 achieves top-tier performance in language and multimodal tasks.
- Open-source models and datasets aim to accelerate AI adoption across industries.
- Specialized models for coding and math set new benchmarks in their domains.
