July 27, 2012 - At yesterday's 2025 World Artificial Intelligence ConferenceAt the "AI Renewal Industry Win-Win" Enterprise Artificial Intelligence Industry Development Forum.China MobileRenewed Release "the nine heavens"Basic Large Model 3.0.
The "Jiu Tian" Basic Model 3.0 released this time is the new achievement of the State-owned Assets Supervision and Administration Commission of the State Council (SASAC)'s "Hundred Projects", which further strengthens the four major features of "high security, high controllability, national production, and industry-wide", focusing on realizing three major revitalizations. It further strengthens the four characteristics of "high security, high controllability, national production, and industry-wide", and focuses on realizing three major innovations:
First, the big model of JiuTianZhongKing language realizes the double leap of architecture and performance.
The model breakthrough adopts the MoE architecture that can be scaled up to trillions, strengthens the reasoning capability with 15T token multi-stage rationed pre-training data and the full-process governance system, and innovatively constructs a two-dimensional hierarchical post-training framework with 113 domains × 53 capabilities, which, combined with the dynamic reinforcement learning strategy, drives the complex reasoning capability to increase by 35%.
In the GPQA-Diamond evaluation, "Jiu Tian" was ranked second in the world with 77.67 points.Beyond DeepSeekR1 and Qwen3In ArenaHard V0.1, the model ranked first in the world with a score of 67.2, and in the BFCL V3 evaluation, it scored 68. While the performance jumps, the model further strengthens the controllable generation ability, cracks the immersive character interpretation, precise process built-in technical details, and realizes the professional scene with zero illusion.
Based on the latest language big model, China Mobile also launched several special models this time. The Nine Days Code Big Model uses a two-stage continuous training technique to support a variety of tasks such as code generation, annotation generation, unit test generation, and code intelligent Q&A, and supports 10 + mainstream programming languages such as Python, Java, JS, TS, Go, C++, and so on.
In terms of metrics, it leads in many code generation lists such as EvalPlus, MHPP, LivecodeBenchv6, and so on. The Nine Days Math model has reached the SOTA level in the industry in both short and long thinking modes, and many of the indexes have surpassed Qwen 2.5 Math, Qwen3, DeepSeek Math, DeepSeek R1-Distill and other models with the same parameter level.
Second, the Jiu Tian Shan Zhi multimodal large model comprehensively improves the generation of controllability and comprehension.
The model is based on innovations such as complex spatio-temporal modeling, progressive joint training of stream-matched images and videos, and end-to-end locally controllable attention mechanism, and at the same time introduces multimodal comprehension information to improve the perception of textual commands and input-conditional images and videos, and joint training of graphic and textual interwoven data to realize the ability to generate high-quality images and videos as well as to perform multilaterally conversational highly controllable and precise editing operations, which significantly enhances the visual generation's Flexibility and Convenience.
For example, in terms of image generation function can support multiple rounds of accurate local modification, such as modifying text, modifying background, adding elements, etc..
The model's performance in graph comprehension and video comprehension is also improved across the board.
- In terms of graph comprehension, Jiu-Tian's model scored 82.2, 64.3, and 94.9 in graph comprehension tasks such as MMStar, HallusionBench, and OCRBench, respectively, which is at the leading level in the industry;
- For video comprehension, the Ninth Heaven model outperforms Qwen2-VL and InternVideo2 in both Videomme and MVbench tasks.
Third, the model and core technology open source.
- First, the open source JiuTian Numerical Kids structured data model, open source content including JT-DA-8B model and subsequent evolution version, support for downloading the model weights, fine-tuning code, inference code and so on;
- The second is the open source Jiu Tian mathematical large model, open source content including JT-Math-8B series of models, support for downloading model weights, inference code, technical reports;
- Third, the open source JiuTian code big model, open source content including JT-Coder-8B series of models, support for downloading model weights, reasoning code, technical reports;
- Fourth, the industry's first open source structured data model evaluation data and TReB evaluation system, the evaluation system covers 6 major tasks, 34 capabilities, including high-quality and comprehensive data, reasoning models and evaluation indicators, support for the download of evaluation datasets, test code;
- Fifth, the open-source CCR-Bench Industry Scenario Complex Instruction Compliance Evaluation dataset, which covers 174 high-quality, diversified, and difficult complex instruction data, highly simulates typical industrial scenarios such as health experts, intelligent customer service, medical assistants, and so on, and supports downloading the dataset.
The above is live on the Huanxin Community, Github, HuggingFace, Magic Hitch Community, Gitee, and Arxiv.
1AI learned from the forum that China Mobile's Jiu Tian Artificial Intelligence Research Institute plans to A technical deep dive into the content of this release was launched in August.