China Telecom releases Xing4.0: an open bet on Huawei chips
China Telecom AI, the artificial intelligence company of the China Telecom group, has officially released Xing4.0-29B-A4B, an open-weight language model with a mixture-of-experts architecture designed for agentic tasks such as tool calling, terminal work and programming. Released under the Apache 2.0 license, the system belongs to the Xing series, formerly known as TeleChat, and has 29 billion total parameters with 4 billion active per token, plus a native 256K context window that can be extended to 512K. According to AIbase, the announcement in China dates back to September 17; the international rollout followed with a press release issued from Beijing on September 22, 2026.
The most interesting aspect is the claim about the supply chain. In the official model card published on Hugging Face, the company states that training took place entirely on Huawei Ascend NPUs using the MindSpore framework, presenting it as “the first model of this scale trained entirely on the Ascend NPU platform”. In the same document, the company claims a training-efficiency gain of about 96%, without specifying the baseline used for the calculation. On deployment, the commercial press release says the model needs 15 GB of memory on a single consumer graphics card, without stating which quantization the figure refers to; AIbase's analysis attributes it to 4-bit quantization, while the Hugging Face card lists no memory requirements and the files on the official repository are currently available in F32 and BF16 formats.
The performance figures released by the developer show mixed results against direct competitors. In the agentic and coding tests cited in the technical card, Xing4.0 beats Gemma4-26B-A4B and Qwen3.6-35B-A3B on Terminal-Bench 2.1, reaching 57.50 points, and on Claw-Eval with 76.55, but falls short of Qwen3.6 on SWE-bench Verified, scoring 75.00 against 76.00 for Alibaba's model. In the same table it also trails Qwen3.6 on AIME2026, Tau3-Bench, SWE-bench Multilingual and AA.LCR, and Gemma4 on IFBench and AA.LCR. The company says it has already integrated the system into its own customer service to handle complex requests, claiming better first-contact resolution but offering no numbers to back it up. On the project's direction, a China Telecom AI spokesperson said: “We believe the future of AI lies not in ever-larger models, but in making powerful intelligence accessible to everyone, affordable and deployable anywhere.”
Performance comparisons published on a company technical card remain provisional until they are independently verified. On the industrial side, however, the claim of training carried out entirely on Huawei hardware is what will matter, and it is also the one nobody outside can check: the weights can be downloaded, the hardware that produced them cannot.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the official model card on Hugging Face (XingChen-AGI organization): architecture, Apache 2.0 license, context, training platform and the full benchmark table with the comparison models. In the official GlobeNewswire press release of 22/09/2026 I checked the date, place, spokesperson quote, the 15 GB requirement and the customer-service use. Independent confirmation: AIbase (parameters, context, 4-bit quantization, SuperCLUE). I could not read the HPCwire/AIwire article (403 error) and did not open the GitHub repository.
- Incertezze
- All benchmarks are self-reported by China Telecom AI and there is no independent verification yet, apart from the SuperCLUE score reported by AIbase and not checked against the primary source. In coding the model does not always beat Qwen3.6-35B-A3B: on SWE-bench Verified it scores 75 against 76, and the press release quotes the 75 without mentioning this. According to AIbase, the 15 GB figure applies only with 4-bit quantization: the press release does not say so and Hugging Face has no quantized variants. The +96% efficiency claim has no stated baseline. Training entirely on Ascend is a company claim that cannot be verified from outside; the customer-service improvement is stated without numbers. The announcement date in China (September 17, according to AIbase) differs from that of the international press release (September 22).
- Perché pubblicarla
- It is an open-source model with a permissive license that can run locally, and the company presents it as trained entirely on Chinese hardware: a relevant topic for technological sovereignty and for the effect of export controls. It is also a chance for an anti-hype piece, because putting the full table next to the press release shows the model does not win everywhere. It does not overlap with topics already published.