What it is
Xing4.0-29B-A4B is an open-weight model from China Telecom Artificial Intelligence Technology, published on September 17, 2026. Xing is the new name for the company's TeleChat series. It is a mixture-of-experts text model with 29 billion parameters, of which 4 billion are active per token, routed over 64 experts plus one shared expert. The native context is 256K tokens, extendable to 512K, and thinking can be switched on or off per request.
China Telecom says it is the first model of this size trained entirely on Huawei's Ascend 910C NPUs with the MindSpore framework, and that its engineering work raised training throughput by about 96% over the stack's out-of-the-box performance. The model is tuned for coding and agent work and has been adapted for harnesses including OpenCode, Claude Code, OpenClaw and Hermes.
The weights are under Apache 2.0 on Hugging Face, ModelScope and Modelers. vLLM and SGLang need out-of-tree plugins from China Telecom; KTransformers, llama.cpp and MLX ports also exist, and OpenRouter lists a free hosted endpoint run by Phala.
Our review
Xing4.0 matters more as an industrial milestone than as a product. Training a competitive 29B model end to end on Ascend 910C clusters shows Chinese labs can build useful models without NVIDIA, and China Telecom published the fused operators and MindFormers recipes to repeat the work.
As a model, it lands roughly level with the comparison it chose. China Telecom's table puts it at 75.0 on SWE-bench Verified against 76.0 for Qwen3.6-35B-A3B and 53.0 for Gemma 4 26B-A4B, and ahead on Terminal-Bench 2.1 (57.5 against 51.5 and 30.0) and Claw-Eval (76.55 against 74.54). Qwen leads on AIME 2026, Tau3-Bench and SWE-bench Multilingual, and Gemma on IFBench and long-context reasoning. These are the vendor's own runs; we found no independent leaderboard entry yet.
The competition has moved on. Alibaba's dense Qwen3.8-27B reports 73.0 on Terminal-Bench 2.1 in its own card, well above Xing's 57.5, and reads images where Xing is text only. Xing's counter is speed: with 4 billion active parameters it does far less work per token than a dense 27B model, which matters for agent loops that burn through long contexts.
The practical rough edges are in serving. Stock vLLM and SGLang do not recognise the architecture without China Telecom's plugins, and the free OpenRouter endpoint is a single third-party host. Community adoption is modest, about 15,000 Hugging Face downloads in the past month, though the card's list of domain fine-tunes, from finance to meeting minutes, suggests the Apache licence is doing its job.
Where it earns its place
An infrastructure team at a Chinese state-owned enterprise on Huawei Ascend hardware
Run a coding and agent model on domestic chips without porting work
Xing4.0 was trained and tuned on Ascend 910C with MindSpore, so the deployment path is documented rather than improvised. It scores close to Qwen3.6-35B-A3B on the vendor's coding benchmarks, under an Apache licence.
A platform engineer serving an internal coding agent on one H100
Run long agent sessions cheaply at 200K-token contexts
The BF16 weights are about 62GB, so one 80GB card holds them with room for cache, and only 4B parameters fire per token. The trade-off is roughly 15 points lower on Terminal-Bench 2.1 than Qwen3.8-27B, a dense model that is slower per token.
A legal-tech start-up with a few thousand labelled contracts
Fine-tune an Apache-licensed base for clause extraction and contract review
China Telecom names contract auditing and table understanding as target fine-tuning domains, and LLaMA-Factory, ms-swift and Unsloth support it. Evaluation is on the start-up: no independent scores exist.
Before you commit
- Setup
- Download from Hugging Face, install China Telecom's xing4 vLLM or SGLang plugin, and serve an OpenAI-compatible endpoint; thinking is toggled with enable_thinking in the chat template.
- Integrations
- vLLM and SGLang (via plugins), KTransformers, llama.cpp GGUF, MLX ports, LLaMA-Factory, MindFormers, ms-swift, Unsloth; adapted for OpenCode, Claude Code, OpenClaw and Hermes. Free endpoint on OpenRouter.
- Data and privacy
- Apache 2.0: commercial use and fine-tuning allowed with no revenue cap or acceptable-use addendum. Self-hosted, nothing leaves your servers; the OpenRouter endpoint is run by a third party, Phala, under its own terms.
- Cost at scale
- BF16 weights are about 62GB (31.2B parameters including embeddings), one 80GB H100; FP8 halves that; a community GGUF build at 11.6GB fits a 16GB card. The only hosted option, OpenRouter's free endpoint, has daily request caps, so production traffic means self-hosting.
- Limits
- Text only; 256K native context; benchmarks self-reported with recommended temperature 0.8 to 1.0; the free hosted endpoint caps completions at about 236K tokens and depends on one provider.
What's good
- Apache 2.0 licence
- 4 billion active parameters per token, cheap to serve for its size
- 256K native context, 512K extended
- Runs on Ascend NPUs as well as NVIDIA, AMD and Apple stacks
What's not
- Text only
- Trails Qwen3.8-27B on Terminal-Bench 2.1 by 15.5 points, per each vendor
- vLLM and SGLang need out-of-tree plugins
- All benchmarks self-reported; no independent leaderboard entry yet
Best for
Teams on Huawei Ascend hardware, or those wanting a small Apache-licensed agent model to fine-tune for domain tasks.
Alternatives
Xing4.0-29B-A4B: common questions
- Is Xing4.0-29B-A4B worth it?
- Xing4.0 is proof that China can train a useful agent model without NVIDIA, and a reasonable cheap coder, but not a reason to leave Qwen. We rate it 3.6 out of 5.
- How much does Xing4.0-29B-A4B cost?
- Free weights under Apache 2.0. No first-party API; OpenRouter lists a free endpoint (hosted by Phala) under its free-model request caps.
- What are the downsides of Xing4.0-29B-A4B?
- Text only; trails Qwen3.8-27B on Terminal-Bench 2.1 by 15.5 points, per each vendor; vLLM and SGLang need out-of-tree plugins; all benchmarks self-reported; no independent leaderboard entry yet.
- Who is Xing4.0-29B-A4B best for?
- Teams on Huawei Ascend hardware, or those wanting a small Apache-licensed agent model to fine-tune for domain tasks.
- How much does Xing4.0-29B-A4B cost for a team or in production?
- BF16 weights are about 62GB (31.2B parameters including embeddings), one 80GB H100; FP8 halves that; a community GGUF build at 11.6GB fits a 16GB card. The only hosted option, OpenRouter's free endpoint, has daily request caps, so production traffic means self-hosting.
- Is Xing4.0-29B-A4B safe for company data?
- Apache 2.0: commercial use and fine-tuning allowed with no revenue cap or acceptable-use addendum. Self-hosted, nothing leaves your servers; the OpenRouter endpoint is run by a third party, Phala, under its own terms.
- What are the best Xing4.0-29B-A4B alternatives?
- The closest alternatives we have reviewed are Qwen3.8-27B, Gemma 4 and gpt-oss-20b.
Answers checked Oct 22, 2026.