Xing 4
China Telecom's Xingchen AGI Lab technical report on a frontier agentic model family trained entirely on Huawei Ascend chips — a 285B-parameter flagship pretrained on 8,192 Ascend 910C cards with the TeleGym agent-RL platform, plus an open-weight 29B-A4B model.
Why it matters
A worked, documented example of a 285B frontier-scale pretraining run with zero Nvidia hardware or CUDA — MindSpore HyperParallel throughout — proving the domestic Chinese training stack now sustains frontier-scale agentic runs.
What you could build with it
A research team building sovereign or sanctions-resilient AI infrastructure could use this report as a blueprint for training large agentic models on non-Nvidia accelerators, and an indie developer could fine-tune the open 29B-A4B model for long-context agent work on a single consumer GPU.
Does it hold up?
Too early to judge — this is a technical report, not a product launch; the 285B flagship has no public weights, and all benchmarks are the authors' own with no independent verification or developer usage reported yet.
Built with Xing 4
- TeleChat Xing4 Trains 285B-Param Agentic MoE Entirely on 8,192 Ascend 910C NPUs With MindSporeaiweekly.co · Summary of the paper: 29B-A4B open model with 256K-1M context, 285B-A17B flagship, MLA plus DeepSeek Sparse Attention, and DSpark speculative decoding claiming 2.67-3.77x throughput.
- 18 Nov 2026 · the stack beneath the tokensweightsandtakes.com · Daily digest noting Xing 4 topped Hugging Face Daily Papers with 146 upvotes and framing it as agent and kernel work moving onto Ascend hardware.
- China Telecom Details Xing4 285B MoE Model Trained on 8,192 Ascend NPUsaf.net · Report on the paper's two models and architecture, noting the 285B flagship and the open-weight 29B-A4B variant.
Learn more
First spotted on article: source.
More AI models
Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73Qwen4-Max: Alibaba's new flagship model goes liveAlibaba's 2.4T-parameter flagship is now generally available via API, with a 1M-token context, native vision, and…models · JEV 0.67smolmaxxing: run 16B LLMs on a 1 GB TV box or a 2013 iPadAn MIT-licensed llama.cpp patch set that runs language models up to a 16B MoE on devices with about 1 GB of usable RAM…models · JEV 0.65Blue TTS v3: a 388MB open text-to-speech model with voice cloningAn open-source 388MB text-to-speech model (MIT) with voice cloning from about 10 seconds of speech, 7 languages, 15…models · JEV 0.65Haru-4BOpen full-duplex spoken-dialogue foundation model that listens and speaks simultaneously (7.5 Hz frames, 133 ms per…models · JEV 0.63
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.