AI Radar Xing4.0-29B-A4B: an open Chinese MoE, 29B heavy, 4B awake

On 16/17 September 2026, XingChen-AGI uploaded Xing4.0-29B-A4B: 29 billion parameters, four active per token, 256K native context, Apache 2.0. The org name stands for the model line of China Telecom's AI research, previously known as TeleChat; the training reports carry the same author groups. Five days after release the card counts 18,394 downloads and 1,177 likes, and the official GGUF repository followed on 20 September. A serious candidate for our test list. With this particular model, though, the supply-chain check belongs on the table, and we run it in the open below.

The technique in one table

The Ascend line is the real message between the lines. This model was trained entirely on Huawei's NPU platform, with custom fused operators; 96 percent more training throughput than the platform's out-of-the-box state, per the vendor. The Western path (NVIDIA GPUs, CUDA, PyTorch) is no longer the only route on which a model of this class comes into existence. Operationally, nothing changes on this side of the map: the weights ship in Transformers format, vLLM, SGLang and KTransformers appear on the card, plus a llama.cpp guide. The card recommends sampling temperature 0.7 for thinking tasks and 0.1 for factual ones. That is detail work, but details like that are how you tell whether a card was written by people who actually ran the model. Here, it reads that way.

What the vendor submits

The card pits the model against two open rivals in the same weight class, measured with documented harnesses, but by the manufacturer itself:

The profile stands out more than the peak. A 4B-active MoE that nearly matches a 3B-active MoE on SWE-bench and clearly beats it on Terminal-Bench is demonstrably aimed at agent work; the card names tool calling, multi-step planning and fine-tuning for contract review and table understanding as target use. Precisely the field where we run our own AI pipeline. A model in this league that reaches 18,000 downloads in a week belongs on the radar, even if it fails our test in the end.

Two abbreviations from the layout row deserve a sentence each. MTP, multi-token prediction, trains the model to predict several tokens ahead; at inference that pays off when a speculative decoder rides along on those predictions. Whether and how the card uses the feature under benchmark stays unclear in the details; the footnotes name temperature, modality and repetition counts, no decoding parameter. mHC stands for hyperconnected residual paths, architectural fine-tuning from the research orbit, without an independently verifiable effect claim. Both are things we handle in the test ourselves rather than judge from the table.

What it costs in hardware

The official GGUF tier is IQ4_NL at 20.1 GB by the GGUF repository's LFS metadata; the card rounds to about 18 GB. The intent is a single consumer GPU. For our cards that means:

• RTX 5090 (32 GB): weights plus MLA KV cache fit with headroom. 256K context stays a promise we have to re-compute; MLA compresses the KV cache heavily, but MLA support in llama.cpp is younger than the model.
• 24 GB cards: the 20 GB file busts the budget together with the context. Expert offload into RAM, the KTransformers and cpu-trick pattern from the scene, is the escape hatch, at the known price: PCIe becomes the bottleneck.
• Our own lane: for 4× RTX 5090 the FP8 variant in the hub is the more direct candidate; its on-disk size is not yet verified by us, and it goes on the bench before any word about deployment.

The lineage, briefly: TeleChat 2.5 and TeleChat 3 came with training reports disclosing data mixtures and infrastructure, published by a state group. That is more provenance than several Western cards this week offer, and at the same time no reason to take any number on faith. Together it yields no advance trust, but a solid basis for testing.

One more reading of the week, which the briefing did not include: the hub's trending list this week showed mostly evergreens, gte-reranker, Qwen3-8B, e5 embeddings, plus quantization mirrors of existing models (the GLM-4.7-Flash GGUF mirrors, most recently refreshed by mradermacher on 17 September). In weeks like this, a genuinely new model is measured against 18k downloads; the rest of the movement is mirror demand for the proven. Xing4.0 is the only substance-new model in that picture.

Supply chain, stated plainly

China Telecom is a state-controlled group. What that does and does not mean:

• License: Apache 2.0. Self-hosting needs no more rights than that; the weights themselves are legally clean.
• Data sovereignty: as long as the model computes in our own network on our own hardware, not one byte travels to China. That is the same argument as for every on-prem model, and the reason we do not run the API variants of international providers at all.
• Residual risk: training-data provenance, and with it copyright questions, are documented no better or worse for this model than for the competition. Verifiable provenance exists for none of this week's open weight releases, Western ones included.
• Maintenance: the GitHub repository is five days old, 82 stars, one open issue. Whether the line survives beyond China Telecom's own projects will be decided over quarters, not days.

Our procedure therefore stays as always: an isolated test environment, our 34-task sample from the self-hosting post, a comparison against the running Qwen3.8-27B lane. Only then a word about deployment. Anyone who sees product recommendations for this model before that test ran should remember the first week of TeleChat. The line is young, and the card is well written. Neither is a quality guarantee, and together they are still a good starting point.

Further reading

• Model card XingChen-AGI/Xing4.0-29B-A4B: architecture, tables, footnotes on the runs
• GGUF repository with IQ4_NL: the official single-GPU build with llama.cpp guide
• GitHub XingChen-AGI/Xing4.0-29B-A4B: serving instructions (vLLM, SGLang, KTransformers)
• Training Report of TeleChat3-MoE (arXiv): predecessor line, same author group
• FP8 variant in the hub: for cards with real memory
• From this site: MoE at the edge, GLM 5.2 in the radar