Most people meet large models through a chat box. China Mobile is betting that the more valuable meeting happens where there is no chat box at all — inside a power grid, a bank's back office, or a government service desk. The country's largest telecom operator thinks the model should live there, in the infrastructure, rather than on a consumer's home screen.
In July 2025, at the World AI Conference in Shanghai, China Mobile released Jiutian (九天) Foundation Model 3.0. It is presented not as a product to download but as a state-owned, "domestic and controllable" foundation model (基础模型) for industry — a base layer the operator can drop into the sectors it already serves.
Key takeaways
- China Mobile released Jiutian (九天) Foundation Model 3.0 in July 2025 at the World AI Conference in Shanghai.
- The language model uses a Mixture-of-Experts (MoE) architecture scalable to trillion-parameter scale, pre-trained on 15 trillion tokens.
- China Mobile reports leading scores on GPQA-Diamond (77.67), ArenaHard (≈67.2), and BFCL V3 (68), and open-sourced several specialist models.
- It is positioned as a state-owned, fully domestic model stack for industry, telecom, and government use rather than a consumer app.
What Jiutian 3.0 is
Jiutian 3.0 is a family, not a single model. The headline language model, "Zhongqing," adopts a MoE architecture that China Mobile says can be extended toward a trillion parameters. It was pre-trained on 15 trillion tokens under a multi-stage data mix, and the company says complex-reasoning ability improved by about 35% through a dynamic reinforcement-learning strategy layered on top of a two-dimensional post-training framework covering 113 domains and 53 capabilities.
Alongside it, China Mobile shipped specialist models for code and mathematics, and a multimodal model, "Shanzhi," focused on controllable image and video generation and understanding. The pitch is "high security, high controllability, full domestic supply chain, full industry coverage" — wording that matters for state-owned buyers who must answer procurement and data-sovereignty questions.
The benchmark story, and how to read it
China Mobile published an aggressive set of benchmark numbers: 77.67 on GPQA-Diamond, which it says places the model second globally; about 67.2 on ArenaHard, which it claims as first globally; and 68 on BFCL V3 for tool-calling. These figures come from China Mobile's own WAIC disclosures and were also reported by Chinese financial media.
Read them with the usual caution. Company-reported benchmark scores are a starting point, not proof. They tell you where the team aimed — reasoning, agentic tool use, and professional-domain accuracy — more than they tell you how the model behaves in a messy production system. The more durable signal is the open-sourcing of the specialist models and evaluation sets, which lets outside engineers verify claims instead of taking them on faith.
Open-source, with a state-owned twist
China Mobile open-sourced several pieces: a structured-data model (JT-DA-8B), a math model (JT-Math-8B), a code model (JT-Coder-8B), and two evaluation frameworks, TReB and CCR-Bench. They are posted on the "焕新社区" (renewal community), GitHub, Hugging Face, ModelScope (魔搭), Gitee, and arXiv.
That combination is unusual. A state-owned enterprise publishing weights and benchmarks on global platforms signals a different theory of value than a closed lab: the operator wins by becoming the default substrate for industry AI, not by metering API calls. It also spun off a dedicated AI research subsidiary, 中移九天人工智能科技, to run the work as a standing business.
Why a telecom operator builds models
China Mobile controls the networks and data centers that other companies rent. A self-built foundation model lets it sell "AI +" as a bundled capability — training, hosting, and inference — to sectors that distrust sending data to a third party. The model is the hook for a larger cloud-and-connectivity play, and the "domestic and controllable" framing is precisely what public-sector and state-owned customers need to hear.
For global readers, the takeaway is that China's model race is not only Big Tech versus startups. The operators, with their balance sheets and captive enterprise demand, are becoming their own model houses.
Honest limitations
Key benchmark figures (GPQA-Diamond 77.67, ArenaHard ≈67.2, BFCL V3 68) originate from China Mobile's own disclosure and Chinese financial-media reports; we did not reproduce the evaluations. The "+35% reasoning" claim is the company's and depends on its internal baseline. We did not test the open-source 8B specialist models, and their real-world quality at that size may differ from the flagship's headline numbers. Licensing terms on the open releases were not independently reviewed here.
What readers can do now
- If you work in enterprise AI, download the JT-Math-8B or JT-Coder-8B releases from Hugging Face or ModelScope and test them on your domain before assuming the flagship's scores transfer.
- If you follow China's AI industry, watch whether operators like China Mobile convert open models into paid "AI +" cloud bundles — that is where the business model lives.
- If you benchmark models, run the open TReB and CCR-Bench suites on competitors too, since China Mobile published them as shared yardsticks rather than internal secrets.