Xing4.0-29B-A4B: China Telecom's Agent-Focused MoE Model, Explained
Xing4.0-29B-A4B is China Telecom's new 29B MoE model trained on Ascend NPUs, benchmarked against Gemma4-26B-A4B and Qwen3.6-35B-A3B.
What is Xing4.0-29B-A4B?
Xing4.0-29B-A4B is an open-weight mixture-of-experts (MoE) language model released by China Telecom Artificial Intelligence Technology Co., Ltd. It’s the latest entry in the Xing series, formerly known as TeleChat, and it’s built specifically for agentic work: tool calling, multi-step planning, coding, and terminal-based tasks. The model has 29 billion total parameters but only activates 4 billion per token, and it’s notable for being trained entirely on Huawei’s Ascend NPU platform rather than Nvidia GPUs.
TL;DR
- Xing4.0-29B-A4B is a 29B-parameter MoE model with only 4B active parameters per token, released by China Telecom under an Apache 2.0 license.
- The model natively supports a 256K context window, extensible to 512K, and uses an mHC + MLA + MTP architecture designed for long-context agent tasks.
- It’s described as the first model of this scale trained entirely on Ascend NPU hardware using the MindSpore/MindFormers stack, rather than the usual Nvidia GPU pipeline.
- China Telecom reports a roughly 96% training throughput improvement over out-of-the-box performance through MoE communication tuning, selective recomputation, and custom fused operators.
- On the model card’s own benchmarks, Xing4.0-29B-A4B leads on Terminal-Bench 2.1, Claw-Eval, and DeepresearchBII against Gemma4-26B-A4B and Qwen3.6-35B-A3B, but trails Qwen3.6 slightly on SWE-bench Verified and SWE-bench Multilingual.
- The model supports deployment through vLLM, SGLang, and KTransformers, and integrates with agent frameworks including OpenCode, Claude Code, OpenClaw, and Hermes.
- Fine-tuning is supported via LLaMA-Factory and MindFormers, positioning the model for domain-specific customization on tasks like contract auditing and table understanding.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
What’s under the hood architecturally?
Xing4.0-29B-A4B runs on a combination the model card calls mHC + MLA + MTP. MLA (multi-head latent attention) is the same attention mechanism family popularized by DeepSeek’s V2/V3 models, designed to compress the key-value cache and cut memory overhead during inference, which matters a lot when you’re pushing context windows out to 256K or 512K tokens. MTP refers to multi-token prediction, a training technique that lets the model predict several tokens ahead per step, generally improving both training efficiency and inference speed. mHC is the piece China Telecom highlights as requiring custom engineering work, including dedicated fused operators built for Ascend hardware.
Structurally, the model has 40 layers, a hidden size of 3584, and a MoE layout with 64 routed experts plus 1 shared expert, of which 4 experts activate per token. That 4B-active-parameter footprint is what keeps inference costs down relative to a dense 29B model, while still giving the router access to a large pool of specialized experts.
Why does training on Ascend NPUs matter?
Most large language models released globally are trained on Nvidia GPU clusters. Xing4.0-29B-A4B is pitched as the first model at this scale trained end-to-end on Huawei’s Ascend 910C NPUs using the MindSpore and MindFormers software stack instead. That’s a meaningful engineering claim: training frameworks, communication libraries, and low-level kernels built around Nvidia’s CUDA ecosystem don’t transfer directly to Ascend silicon, so getting a MoE model of this size to train efficiently requires rebuilding a lot of the pipeline from scratch.
China Telecom’s team reports several specific optimizations that got them there: fine-grained MoE communication tuning (important because MoE routing creates heavy all-to-all traffic between devices), selective recomputation to manage memory, DVM automatic graph-operator fusion, and custom Ascend C operators for the mHC architecture piece. Combined, the model card states these changes improved training throughput by approximately 96% over the unoptimized baseline. That’s a training-infrastructure story as much as a model-capability story, and it signals a maturing alternative to the Nvidia-dependent training stack that most labs still rely on.
How does it perform on agent and coding benchmarks?
The model card runs Xing4.0-29B-A4B against two comparably sized MoE models: Gemma4-26B-A4B and Qwen3.6-35B-A3B. All three are in the same rough weight class, with similarly small active-parameter counts, so the comparisons are apples-to-apples on inference cost even though total parameter counts differ slightly.
On the benchmarks that matter most for the agentic use case China Telecom is targeting, Xing4.0-29B-A4B does well:
- Terminal-Bench 2.1: 57.50, well ahead of Qwen3.6 (51.50) and far ahead of Gemma4 (30.00).
- Claw-Eval: 76.55, the highest of the three, versus 74.54 for Qwen3.6 and 71.49 for Gemma4.
- DeepresearchBII: 60.80, ahead of Qwen3.6 (59.70) and well clear of Gemma4 (39.30).
These three benchmarks specifically test agentic behavior: running commands in a terminal environment, tool-use evaluation, and multi-step research tasks with web/MCP tool access. That’s consistent with the model’s positioning as an “agent-oriented” release rather than a general chatbot.
On raw coding benchmarks the picture is more mixed. Qwen3.6-35B-A3B edges out Xing4.0 on both SWE-bench Verified (76.00 vs 75.00) and SWE-bench Multilingual (67.20 vs 66.00), though the gap is small. On instruction-following (IFBench) and long-context reasoning (AA.LCR), Gemma4 comes out slightly ahead of Xing4.0. Xing4.0 posts the strongest AIME2026 math score of the three (90.00), though Qwen3.6 is close behind at 92.70 and actually edges it out there. Tau3-Bench, a tool-use benchmark, favors Qwen3.6 (67.20) over Xing4.0 (64.63) and Gemma4 (58.90).
The overall pattern: Xing4.0-29B-A4B isn’t uniformly the strongest model in this trio, but it’s specifically strong on terminal and long-horizon agent tasks, which lines up with the architectural choices (long context, mHC, tool-calling focus) the model card emphasizes.
Is Xing4.0-29B-A4B worth using for agent workflows?
For teams building coding agents or terminal-automation tools, the benchmark profile is worth taking seriously. Terminal-Bench 2.1 and Claw-Eval scores that lead a same-class Qwen model suggest real strength in the specific niche China Telecom optimized for, and the 256K native context (with 512K extension) gives it headroom for long agent traces without needing aggressive context management. The active parameter count (4B) also keeps inference costs comparable to much smaller dense models, which matters for anyone running agent loops with many tool calls per task.
The tradeoffs are on general reasoning and instruction-following, where Gemma4 and Qwen3.6 each have an edge in specific spots. For teams whose workloads lean toward SWE-bench-style code editing over terminal or research-agent tasks, Qwen3.6-35B-A3B’s numbers are marginally better on those specific benchmarks. As with any newly released model, none of these figures come from a third-party leaderboard, they’re self-reported by the model’s own developer using their own evaluation harnesses and hyperparameters, so independent verification is worth doing before making a production decision.
Practically, the model supports the common serving stacks (vLLM, SGLang, KTransformers) and fine-tuning frameworks (LLaMA-Factory, MindFormers), and it’s released under an Apache 2.0 license, which removes licensing friction for commercial use. That combination, open license plus mainstream framework support plus agent-framework integrations (OpenCode, Claude Code, OpenClaw, Hermes), makes it straightforward to try in an existing agent stack without deep infrastructure changes.
Frequently Asked Questions
What does “A4B” mean in Xing4.0-29B-A4B?
It refers to “4B active,” meaning that while the model has 29 billion total parameters spread across its mixture-of-experts layers, only about 4 billion parameters are actually used to process each token, since just 4 of the 64 routed experts (plus 1 shared expert) fire per forward pass.
How is Xing4.0-29B-A4B different from previous TeleChat models?
Xing4.0-29B-A4B is the successor to the TeleChat3 line and the latest release in what China Telecom now calls the Xing series. Its cited technical report is titled “Training Report of TeleChat3-MoE,” confirming the lineage, while the model card frames Xing4.0 as a next-generation, agent-oriented iteration with a new architecture (mHC + MLA + MTP) and native Ascend NPU training.
Can I fine-tune Xing4.0-29B-A4B on my own data?
Yes. The model card lists support for LLaMA-Factory and MindFormers as fine-tuning frameworks, and specifically calls out lightweight customization for tasks like intent classification, table understanding, contract auditing, and knowledge-based QA as intended use cases.
What hardware do I need to run it?
The model card doesn’t publish specific VRAM requirements, but with 41 safetensors shards and 29B total parameters, it’s sized similarly to other 20-30B class MoE models. It supports vLLM, SGLang, and KTransformers for serving, all of which handle MoE model deployment across single or multi-GPU setups.
Are the benchmark numbers independently verified?
No. All benchmark comparisons against Gemma4-26B-A4B and Qwen3.6-35B-A3B come directly from the Xing4.0-29B-A4B model card, using evaluation harnesses and hyperparameters chosen by China Telecom’s own team. The model card does document specific settings (temperature, context window, number of runs averaged) for each benchmark, which adds transparency, but the results haven’t been reproduced by a neutral third party.
