A telecom operator best known for broadband and mobile plans is now shipping artificial intelligence models — and not toy ones. In mid-September 2026, China Telecom's AI subsidiary released Xing4.0-29B-A4B (星辰大模型), a model built around a deceptively simple bet: the future of useful AI is not the biggest model, but the smallest one that can still do the work.
The work, in agent terms, means planning a task, calling tools, and delivering a result — not just answering a question.
The bet: small model, big job
Xing4.0 uses a Mixture-of-Experts (MoE) architecture with 29 billion total parameters but only about 4 billion active at any moment. That asymmetry is the whole point. By activating a small slice of the network per task, the model stays cheap to run while keeping a large knowledge base in reserve.
China Telecom positions it as the first 100-billion-class (百亿参数) model trained entirely on domestic compute and a domestic framework, optimized for complex engineering and agent tasks. The company says it natively handles a 256K-token context, extendable to 512K, long enough to keep a multi-step job coherent without forgetting the plan halfway through.
Trained on Chinese chips, end to end
The detail that matters most to Beijing — and to anyone tracking supply chains — is the stack. The model was trained on Ascend 910C (昇腾910C) clusters using the MindSpore / MindFormers framework, with custom operators written for the hardware. China Telecom describes the path as "国芯训国模" — national chips training a national model.
The engineering claim is that this was not a port of a foreign recipe onto local silicon, but a co-design: architecture tweaks, compiler and operator work, and training-infrastructure optimization together lifted training throughput by about 96% over the out-of-box baseline, nearly doubling it.
The consumer-GPU math
Here is the part that should surprise people outside China's AI labs. After 4-bit quantization, Xing4.0's memory footprint drops to roughly 15 GB — about 75% less than the ~60 GB an uncompressed 29B-class model would need. That means it runs on a single RTX 3090 or RTX 4090, consumer cards with 24 GB of memory that many developers already own. A new RTX 4090 sells for around RMB 12,000 ≈ US$1,650 / HK$12,900, a world away from the multi-card server farms usually required for serious models.
Reported inference speed is about 30 tokens per second on that hardware — slow by data-center standards, but perfectly usable for a local assistant, a private helpdesk, or an edge device that cannot ship data to the cloud. For a small team, "private by default" is a feature, not a compromise.
What it is actually good at
This is an agent model, so the benchmarks lean toward execution rather than trivia. China Telecom cites:
- SWE-bench Verified: 75.0, close to Qwen3.6-35B-A3B at 76.0 and well ahead of a same-class Gemma model at 53.0 — a coding-agent test.
- Terminal-Bench 2.1: 57.5, ahead of comparison models at 51.5 and 30.0 — autonomous command-line task solving.
- Claw-Eval: 76.55, topping comparison models at 74.54 and 71.49 — general agent capability.
- A SuperCLUE agent-score of 93.52, placing third and within one point of the top two Qwen models.
The framing is deliberate: a smaller model that "punches up" against models two to three times its active size on agent tasks, not on generic knowledge quizzes.
Open, and already inside the call center
True to an open-ecosystem stance, the weights are published on GitHub, Hugging Face, Gitee, ModelScope (魔搭), and Modelers (魔乐), with compatibility for mainstream inference frameworks and several domestic AI chips beyond Ascend. China Telecom says it already runs privately inside its own operator hotline, where a customer's call data never leaves the domain and the agent autonomously handles plan changes, bill queries, and fault reports along an "intent understanding → task breakdown → tool call → result assembly → compliance check" chain.
The company's broader TeleAgent (星辰超级智能体) claims more than 1.18 million registered users and a skill library grown from a few thousand to over 50,000 — a sign the agent push is moving from prototype to deployment, not just a model-drop press release.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the strategic signal here is less the benchmark scores and more the "full-stack domestic" path — a model trained on Chinese chips, with Chinese frameworks, that a small team can actually run, is the real capability being stress-tested.
Honest limitations
The benchmark comparisons are drawn from China Telecom's own releases and an ecosystem-affiliated evaluation; we found no independent third-party re-test of Xing4.0 at publication time, so treat the scores as vendor-anchored. At ~30 tokens per second on a gaming GPU, this is not a high-throughput serving engine — it suits local, low-concurrency, privacy-sensitive uses, not a public-facing service handling millions of requests. The "fully domestic stack" claim is accurate for training and the cited chips, but real-world deployment still depends on a maturing domestic software ecosystem whose tooling lags the established Western stacks in some areas. And 4-bit quantization trades some accuracy for size; teams with strict output-quality needs should validate on their own tasks before trusting it blindly.
What readers can do now
- Try it locally for free. Pull the weights from Hugging Face or ModelScope and run a private agent on a spare gaming GPU — no cloud bill, no data leaving your machine.
- Prototype one internal task (a report generator, a file sorter) where privacy matters, to see whether a small domestic model clears your quality bar.
- Watch the "domestic stack" trajectory. If a 29B model trained on Ascend can match larger models on agent tasks, the cheap-local-AI thesis deserves a place in your planning.