China Telecom's 29B agent model was trained without Nvidia
China's largest state carrier put an open-weight coding agent on the market today, and the interesting part is not the benchmark row — it is the stack underneath it.
China Telecom's AI arm released Xing4.0-29B-A4B, a 29-billion-parameter mixture-of-experts coding agent that activates 4 billion parameters per token and was trained end to end on Huawei's Ascend NPU platform using the MindSpore framework. The company describes it as the first 100-billion-class Chinese model trained entirely on domestic compute and a domestic framework and optimized for complex engineering tasks — agentic planning, tool calls, terminal work and long-horizon execution. Weights are Apache 2.0 and posted across Hugging Face, ModelScope, Gitee and Modelers, with a 256K context window extensible to 512K, FP8 and GGUF variants already up, and adaptation for existing agent harnesses including OpenCode, Claude Code, OpenClaw and Hermes.
The architecture is a 40-layer design built on mHC, MLA attention and multi-token prediction, with 64 routed experts, four active per token and one shared expert. China Telecom credits multi-level co-optimization on the Ascend 910C clusters — fine-grained MoE communication, selective recomputation, automatic graph-operator fusion, fused Ascend C operators — with roughly a 96% lift in training throughput over out-of-the-box performance. That efficiency claim is the load-bearing one: it is the argument that a frontier-adjacent agent model no longer requires a CUDA supply chain.
On the model card's own evaluations, the release trades wins with the models it names as peers, Gemma4-26B-A4B and Qwen3.6-35B-A3B. It reports 57.50 on Terminal-Bench 2.1 against 30.00 and 51.50, 75.00 on SWE-bench Verified against 53.00 and 76.00, 76.55 on Claw-Eval, 60.80 on DeepresearchBII and 64.63 on Tau3-Bench. Qwen still leads on AIME2026 (92.70 to 90.00) and edges SWE-bench Verified, and Gemma leads on IFBench — so the honest read is a competitive small-agent model, not a new ceiling. These are vendor-run numbers; nobody outside China Telecom has reproduced them yet.
Why it matters is the combination, not the scoreboard. A state carrier shipping an Apache-2.0 agent model trained on Ascend silicon and a domestic framework is a concrete data point on how much of the agent layer can be built off CUDA — and the card explicitly targets third-party harnesses rather than shipping its own. It lands alongside the company's other agent work, which we covered this week — Ant Group rolls Qianwen Office out to every employee.
What to watch: whether the Terminal-Bench and Claw-Eval results survive independent reruns, and whether other labs adopt the Ascend-plus-MindSpore recipe now that one has published the training playbook with it.
Does an agent model trained entirely off CUDA change your deployment math, or is the harness still the deciding layer? Tell us in the comments.
Sources: XingChen-AGI (GitHub)
