On July 26, at the "AI Renewal, Industrial Synergy" Enterprise Artificial Intelligence Industry Development Forum of the 2025 World Artificial Intelligence Conference, China Mobile officially launched its independently developed "Jiutian" Foundation Large Model 3.0. This release not only demonstrates China Mobile's deep self-research capabilities and continuous innovation in large model fields, but also marks a solid step forward in the development of artificial intelligence industrialization.
As an important achievement of the State-owned Assets Supervision and Administration Commission's "Top 100 Projects" in strategic emerging industries, "Jiutian" Foundation Large Model 3.0 has achieved three key "Renewals" on the basis of inheriting the four characteristics of "High Security, High Controllability, Domestic Production, and Full Industry":
First, Jiutian Zhongqing Language Large Model: Dual Advancements in Architecture and Performance
The "Jiutian Zhongqing Language Large Model" has achieved a breakthrough in architecture, adopting a MoE architecture that can scale up to trillions of parameters. With multi-stage proportional pre-training data of 15T tokens and a full-process governance system, its reasoning capability has been significantly enhanced. The model has also innovatively constructed a two-dimensional hierarchical post-training framework of 113 domains × 53 capabilities, combined with dynamic reinforcement learning strategies, which has improved complex reasoning capabilities by 35%.
In authoritative ranking evaluations, the "Jiutian" language large model has achieved remarkable results:
In the GPQA-Diamond evaluation, it scored 77.67, ranking second globally, surpassing DeepSeekR1 and Qwen3.
In ArenaHard V1.0, it scored 67.2, ranking first globally.
In BFCL V3, it reached 68.
At the same time, the model further enhances controllable generation capabilities, achieving zero hallucination in professional scenarios through precise process internalization and other technical details, solving the problem of immersive role performance.
Based on the latest language large model, China Mobile has also launched several specialized models simultaneously:
Jiutian Code Large Model: It adopts a two-stage continuous training technology, supporting tasks such as code generation, comment generation, unit test generation, and intelligent code QA, covering over 10 mainstream programming languages including Python, Java, JS, TS, Go, and C++. It performs well on multiple code generation rankings such as EvalPlus, MHPP, and LivecodeBenchv6.
Jiutian Mathematics Large Model: It reaches SOTA levels in both short and long thinking modes, outperforming models of similar parameter scales such as Qwen2.5Math, Qwen3, DeepSeek Math, and DeepSeek R1-Distill in multiple indicators.
Second, Jiutian Shanzhi Multimodal Large Model: Comprehensive Improvement in Generative Controllability and Understanding Ability
The "Jiutian Shanzhi Multimodal Large Model" introduces innovative technologies such as complex spatiotemporal modeling, flow-matching image and video progressive joint training, and end-to-end local controllable attention mechanisms. At the same time, by integrating multimodal understanding information and joint text-image-video training, the model's perception ability for text instructions and input condition images and videos has been significantly improved. This means the model can not only generate high-quality images and videos but also perform multi-turn conversational high-controllable precise editing operations, greatly enhancing the flexibility and convenience of visual generation. For example, in image generation, it supports multi-turn precise local modifications, such as modifying text, changing the background, or adding elements.
The model's performance in image and video understanding has also been significantly improved:
In image understanding: In tasks such as MMStar, HallusionBench, and OCRBench, the Jiutian model obtained high scores of 82.2, 64.3, and 94.9, respectively, placing it at the forefront of the industry.
In video understanding: It performed leadingly in the Videomme and MVbench tasks, surpassing Qwen2-VL and InternVideo2.
Third, Open-Source Models and Core Technologies to Promote Industrial Ecosystem Development
China Mobile has also adopted an open strategy, open-sourcing multiple models and core technologies to promote industry development:
Open-source Jiutian Structured Data Large Model: Including the JT-DA-8B model and subsequent versions, it supports downloading model weights, fine-tuning code, and inference code.
Open-source Jiutian Mathematics Large Model: Including the JT-Math-8B series models, it supports downloading model weights, inference code, and technical reports.
Open-source Jiutian Code Large Model: Including the JT-Coder-8B series models, it supports downloading model weights, inference code, and technical reports.
Open-source TReB Evaluation System and Structured Data Evaluation Datasets, the First of Its Kind in the Industry: Covering six major tasks and 34 capabilities, including high-quality, comprehensive data, reasoning patterns, and evaluation metrics, it supports downloading evaluation datasets and testing code.
Open-source CCR-Bench Industry Scenario Complex Instruction Following Evaluation Dataset: It includes 174 high-quality, diverse, and high-difficulty complex instruction data, highly simulating typical industrial scenarios such as health experts, intelligent customer service, and medical assistants. It supports downloading the dataset.
All of the above open-source content has been launched on the new community, Github, HuggingFace, Moba Community, Gitee, and Arxiv, providing valuable resources for developers and researchers to jointly accelerate the innovation and application of artificial intelligence technology.