A phone company is not the first place you look for cutting-edge AI research. Yet China Telecom, the country's largest fixed-line carrier, quietly became one of the few organizations to open-source a very large language model trained without a single foreign-made AI chip.
The model is called TeleChat2, and the interesting part is not just its size. It is where it was born.
Key takeaways
- TeleChat2 is a large model (大模型) family from China Telecom's AI institute (TeleAI), branded under the "XingChen" (星辰) name.
- The 115B version, open-sourced in September 2024, was trained on 10 trillion tokens using only domestic computing power and a domestic deep-learning framework.
- China Telecom says it was the first fully homegrown, open-source model at the 100-billion-parameter scale, a claim tied to its use of Chinese chips and software.
- Smaller open releases — 3B, 7B, 35B with tool-calling, plus a 39B-A12B mixture-of-experts — followed through late 2024 and into 2025, all free to download.
The chip question nobody can ignore
Modern AI models are only as free as the hardware that trains them. For years, the most capable open models depended on American GPUs, which face export controls to China. A model that is "open" but can only be rebuilt on banned chips is open with an asterisk. Export controls turn that asterisk into a procurement problem rather than a philosophical one: weights that assume restricted accelerators leave users dependent on hardware they may not be able to buy.
TeleChat2-115B was trained on what China Telecom describes as entirely domestic compute — Chinese-made accelerators — and a domestic deep-learning framework. Ten trillion tokens is what makes the domestic-compute claim expensive to test: at that volume, accelerator supply, framework maturity and cost all meet in the same run. If that holds, it is a meaningful marker: proof that a 100B-plus model can be trained and released without foreign hardware in the loop. The company frames the 115B release as filling a gap in China's experience training models at the very largest scale on homegrown stacks.
A family, not a one-off
The open-source story spans more than the flagship. China Telecom released TeleChat2-3B, 7B, and 35B in November 2024, each with function-calling so the model can actually use external tools — query a database, call an API. A mixture-of-experts version, TeleChat2-39B-A12B, followed in March 2025, a design that activates only part of the model per task to save compute. All four smaller releases are free to download, so a team can test tool-calling and Chinese-language behaviour on its own hardware before committing to the 115B version.
The institute reports that TeleChat2's general capability improved more than 29 percent over the earlier TeleChat1 across reasoning, summarization, long-form writing, and math. That figure comes from the lab's own evaluation set, so treat it as directional rather than independently audited, and note that the comparison is against its own predecessor, not against another vendor's model.
Why a carrier builds models
China Telecom is not chasing chatbots for fun. It runs mountains of customer, network, and operations data, and a sovereign, self-contained model lets it embed AI into services without depending on outside suppliers for either the software or the training stack. For a national infrastructure operator, "we can rebuild this ourselves" is a business continuity plan, not a hobby.
The open-source release doubles as a talent and ecosystem play: universities and smaller firms can build on TeleChat2, and some of that work flows back as adoption and credibility.
What domestic compute actually buys
Training a 100B-plus model end to end on domestic hardware is less about beating a benchmark and more about removing a single point of failure. When the compute, the framework and the weights all sit under one roof, the model stops being a hostage to supply chains. That is an argument about resilience, not a benchmark claim.
Honest limitations
The "first fully homegrown open 100B model" label is China Telecom's own framing and depends on how strictly you define "domestic" across every component, so it should not be taken as a universally accepted record. Performance claims, including the 29 percent gain over TeleChat1, rest on the institute's internal benchmarks rather than a neutral third party. A model trained on domestic chips may also trail the absolute frontier set by labs with the newest American hardware. And open weights are not the same as open training compute — reproducing TeleChat2-115B from scratch would still be a massive, costly undertaking even with the recipe public.
What readers can do now
- If you build for Chinese-regulated industries, test TeleChat2's smaller versions as a sovereign alternative to foreign-hosted models.
- If you follow geopolitics of tech, track whether "trained on domestic compute" becomes a standard selling point for Chinese open models.
- If you are a developer, download the 7B or 35B release and run a local tool-calling agent without relying on overseas cloud APIs.