Xing4.0-29B-A4B Benchmarks: China Telecom's MoE Model Tested
Xing4.0-29B-A4B benchmark scores on SWE-bench, Terminal-Bench, and Claw-Eval, compared against Qwen3.6-35B-A3B and Gemma4-26B-A4B.
What is Xing4.0-29B-A4B?
Xing4.0-29B-A4B is an open-weight mixture-of-experts (MoE) language model from China Telecom Artificial Intelligence Technology Co., Ltd., released under the Xing series name (formerly known as TeleChat). It has 29 billion total parameters but activates only 4 billion per token, using 64 routed experts plus one shared expert, with 4 experts active on any given token. It natively handles a 256K context window, extensible to 512K, and is built on an mHC + MLA + MTP architecture aimed squarely at agentic workloads: tool calling, multi-step planning, and long-horizon reasoning chains.
What makes it notable beyond the spec sheet is where it was trained. Xing4.0-29B-A4B is described as the first model of this scale trained entirely on Huawei’s Ascend 910C NPU platform using the MindSpore/MindFormers framework, rather than on Nvidia GPUs. The model card reports a roughly 96% improvement in training throughput over out-of-the-box performance, attributed to MoE communication tuning, selective recomputation, automatic graph-operator fusion (DVM), and custom Ascend C fused operators for the mHC architecture.
TL;DR
- Xing4.0-29B-A4B is a 29B-parameter MoE model with only 4B active parameters per token, positioning it as a lightweight-inference, heavyweight-capability design.
- It was trained entirely on Ascend 910C NPUs with MindSpore/MindFormers, a departure from the Nvidia-centric training stack most open models rely on.
- On agentic and coding benchmarks it does well: 75.00 on SWE-bench Verified, 57.50 on Terminal-Bench 2.1, and 76.55 on Claw-Eval, all ahead of Gemma4-26B-A4B.
- Against Qwen3.6-35B-A3B, a larger active-parameter competitor, Xing4.0 trades wins: it leads on Terminal-Bench and Claw-Eval but trails slightly on SWE-bench Verified, SWE-bench Multilingual, and Tau3-Bench.
- The model supports 256K native context (512K extensible) and ships with compatibility for vLLM, SGLang, KTransformers, and agent frameworks like OpenCode and Claude Code.
- Its biggest relative weakness in the published numbers is AA.LCR (61.00), where both competing models score higher.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How does Xing4.0-29B-A4B perform on coding benchmarks?
Coding and terminal-use benchmarks are where Xing4.0-29B-A4B makes its strongest case. On SWE-bench Verified, a benchmark that scores a model’s ability to resolve real-world GitHub issues, it scores 75.00. That’s a wide margin over Gemma4-26B-A4B (53.00) and just one point behind Qwen3.6-35B-A3B (76.00), despite Qwen3.6 having roughly 30% more active parameters per token.
On SWE-bench Multilingual, which extends the same issue-resolution task across non-English codebases, Xing4.0 scores 66.00, again well ahead of Gemma4 (51.00) and close to Qwen3.6 (67.20).
Terminal-Bench 2.1 shows the largest gap in Xing4.0’s favor: 57.50 versus Gemma4’s 30.00 and Qwen3.6’s 51.50. This benchmark tests a model’s ability to operate inside a terminal environment across extended sessions (the model card notes evaluations ran with a 24-hour timeout and were averaged over three runs), which lines up with the model’s stated focus on agentic, tool-using tasks over long contexts.
Taken together, the coding numbers suggest Xing4.0-29B-A4B was specifically tuned for engineering and terminal-agent use cases, not just general chat quality.
How does it compare on reasoning and instruction-following?
The picture is more mixed outside of coding. On AIME2026, a competition math benchmark, Xing4.0 scores 90.00, ahead of Gemma4 (88.30) but behind Qwen3.6 (92.70). All three models cluster in a similar range, suggesting comparable raw reasoning capability at this scale.
On IFBench, which measures instruction-following precision, Xing4.0 scores 69.67, sitting between Qwen3.6 (65.50) and Gemma4 (72.67). It’s a middle-of-the-pack result rather than a standout.
AA.LCR is the one benchmark where Xing4.0 trails both competitors, at 61.00 versus Gemma4’s 66.00 and Qwen3.6’s 62.00. The model card doesn’t detail what AA.LCR measures beyond the score itself, so it’s worth treating as one data point among several rather than a definitive weakness.
Tau3-Bench, an agentic tool-use benchmark from Sierra Research evaluated with pass^1 scoring across four runs, has Xing4.0 at 64.63, ahead of Gemma4 (58.90) but behind Qwen3.6 (67.20). This is a case where the larger active-parameter Qwen model pulls ahead on a task that stresses multi-turn tool interaction.
Is Xing4.0-29B-A4B worth using for agentic workflows?
For teams building coding agents or terminal-automation tools, the benchmark profile makes a reasonable case. Xing4.0-29B-A4B leads or nearly matches Qwen3.6-35B-A3B on SWE-bench Verified and SWE-bench Multilingual while using less than a fifth of Qwen3.6’s total parameter count and a comparable active-parameter footprint (4B versus 3B). It clearly outperforms Gemma4-26B-A4B across every coding and agentic benchmark reported, sometimes by wide margins.
Two other benchmarks reinforce the agentic framing. Claw-Eval, evaluated with the model’s official harness across three runs, gives Xing4.0 the top score of the three models at 76.55. DeepresearchBII, an agentic research benchmark run through the OpenCode harness with an Exa MCP server enabled for live search, also favors Xing4.0 at 60.80, well ahead of Gemma4 (39.30) and slightly ahead of Qwen3.6 (59.70).
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The model card also lists explicit compatibility with agent frameworks including OpenCode, Claude Code, OpenClaw, and Hermes, plus fine-tuning support through LLaMA-Factory and MindFormers. That combination, strong terminal and coding scores plus stated framework integration, points toward a model built for deployment inside coding-agent pipelines rather than as a general-purpose chatbot.
Where it’s less convincing is in pure reasoning and instruction-following, where it sits in the middle of the pack rather than leading. Teams whose primary need is math reasoning or strict instruction adherence may find Gemma4 or Qwen3.6 marginally better fits depending on the specific task.
What does the Ascend NPU training background mean in practice?
For most users, the training hardware behind a model is invisible at inference time. Xing4.0-29B-A4B runs through standard OpenAI-compatible APIs and deploys via vLLM, SGLang, or KTransformers like any other open-weight model, so the fact that it trained on Ascend 910C clusters with MindSpore rather than Nvidia GPUs and PyTorch doesn’t change how you’d call or fine-tune it.
What it does signal is a maturing non-Nvidia training stack. The model card’s claim of a 96% throughput improvement from Ascend-specific optimizations, including custom Ascend C operators for the mHC architecture, indicates real engineering investment in making large MoE training viable outside the usual GPU ecosystem. For organizations evaluating vendor diversity in their AI infrastructure, that’s a more relevant signal than for someone just trying to run inference.
Frequently Asked Questions
What does “A4B” mean in Xing4.0-29B-A4B?
It denotes “4B active” parameters. The model has 29 billion total parameters spread across 64 routed experts plus a shared expert, but only about 4 billion parameters are activated for any given token, which keeps inference cost closer to a much smaller dense model.
How does Xing4.0-29B-A4B compare to Qwen3.6-35B-A3B overall?
It’s competitive rather than dominant. Xing4.0 wins on Terminal-Bench 2.1, Claw-Eval, and DeepresearchBII, while Qwen3.6 leads on SWE-bench Verified, SWE-bench Multilingual, Tau3-Bench, and AIME2026. IFBench and AA.LCR split in Qwen3.6’s favor on one and against it on the other.
What context length does Xing4.0-29B-A4B support?
It natively supports 256K tokens of context, with the model card stating this is extensible to 512K. Several benchmark evaluations, including Claw-Eval and DeepresearchBII, were run using the full 256K context window.
Can Xing4.0-29B-A4B be fine-tuned for specific domains?
Yes. The model card lists support for LLaMA-Factory and MindFormers for fine-tuning, and describes the model as suited to lightweight domain adaptation for tasks like intent classification, table understanding, contract auditing, and knowledge-based question answering.
Which inference frameworks support Xing4.0-29B-A4B?
The model is compatible with Hugging Face Transformers, vLLM, SGLang, and KTransformers, and it exposes an OpenAI-compatible API for serving, making it deployable with the same tooling used for other open-weight MoE models.
