China Telecom releases Xing4.0-29B-A4B, a 29B-total/4B-active MoE trained only on Ascend NPUs and MindSpore with 256K context (community report)
Verified (fact)
Xing4.0-29B-A4B, a new model in the Xing series that succeeds TeleChat, has been released, and bartowski's GGUF quantizations are on Hugging Face. Of its 29B total parameters, only 4B are active per token, and it supports a 256K context by default and up to 512K. The developer said it is the first model of this size trained solely on Ascend NPUs and MindSpore. A llama.cpp support PR has also been opened.
Needs fact-check (claim)
The developer claims it is the first model at this scale trained solely on Ascend and MindSpore and that it is optimized for complex engineering tasks.
Why it matters (interpretation)
An open model trained only on Chinese domestic chips has made it into the local inference ecosystem, illustrating the trend of building models without NVIDIA.
Confidence medium · official source pending
Sources
- Community bartowski/XingChen-AGI_Xing4.0-29B-A4B-GGUF · Hugging Face Reddit
More about China Telecom
-
32 export-banned NVIDIA B300 GPUs spotted on China Telecom's network; location and sourcing unconfirmed (via The Information)
The Information reported on October 22 (local time) that an Israeli security firm, while examining exposed monitoring systems, found 32...
-
China Telecom AI lab open-sources Xing4.0-29B-Intent, a dedicated intent-understanding model (report)
The Xingchen general AI lab of China Telecom's artificial intelligence subsidiary announced Xing4.0-29B-Intent, a model specialized for...