Can Xing4.0-29B-A4B really run on one consumer GPU?

With quantization, yes; at full precision, no. China Telecom's claim is that low-bit quantization and memory optimization bring it to roughly 15 GB of GPU memory, which fits a 16 GB card and leaves headroom on a 24 GB RTX 4090 or 3090. Unquantized it needs about 62.9 GB of VRAM per LLM Explorer's calculation, which is a multi-GPU or datacenter configuration. The official GGUF repository is what makes the single-card claim practical — pull a 4-bit quant and run it in llama.cpp, LM Studio or Ollama.

How does 29B total with 4B active change performance?

Mixture-of-experts decouples memory cost from compute cost. You must hold all 29 billion parameters in memory, but only 4 billion participate in each token, so generation speed resembles a 4B dense model while quality tracks something much larger. Practically that means Xing4.0-29B-A4B is fast on modest hardware but still memory-hungry — the opposite trade-off from a dense 29B model, which would be slow and equally memory-hungry. This is why MoE has become the default architecture for locally deployable agent models in 2026.

What is Xing4.0-29B-A4B good at?

China Telecom positions it as an execution model rather than a chat model: multi-step task planning, external tool calling and long-document processing, aimed at engineering work, coding and analysis. It publishes SWE-bench Verified and SWE-bench Multilingual results using the SWE-agent harness at a 210K context window, Terminal-Bench 2.1 using the terminus-2 harness, and Claw-Eval using the official harness at 256K context. Third-party write-ups report coding scores ahead of Gemma 4 and Qwen3.6 on several benchmarks, though independent verification on the Artificial Analysis index was not published as of September 26, 2026.

Why does it matter that it was trained on Ascend NPUs?

It is a proof point that a competitive agentic model can be trained end to end on Huawei's domestic accelerator stack without Nvidia silicon. For Chinese operators under export controls this removes a dependency; for everyone else it signals that the Ascend software stack has matured past the point where it only serves inference. Combined with Huawei's Atlas 960E SuperPod at the cluster level, it makes the Chinese training pipeline a second viable path rather than a fallback.

Quick Answer

What is Xing4.0-29B-A4B, China Telecom's Single-GPU Agent Model?

Q: What is Xing4.0-29B-A4B?

Xing4.0-29B-A4B is an open-weight agentic language model released by China Telecom AI (TeleAI) on September 22, 2026. It is a mixture-of-experts model with 29 billion total parameters and 4 billion activated per token, natively supporting a 256,000-token context extensible to 512,000. It is the latest generation of the TeleChat line, trained entirely on Huawei Ascend NPUs rather than Nvidia hardware, released under Apache 2.0 on the XingChen-AGI GitHub organization and Hugging Face, with FP8 and GGUF variants published alongside the base weights.

Published: September 26, 2026

The short answer

Xing4.0-29B-A4B is China Telecom AI’s open-weight agentic model, released September 22, 2026: 29 billion total parameters, 4 billion active per token, 256K native context extensible to 512K, Apache-2.0 licensed, trained entirely on Huawei Ascend NPUs. China Telecom says low-bit quantization brings it to roughly 15 GB of GPU memory, putting a genuinely agentic model on a single consumer graphics card. FP8 and GGUF builds shipped with the base weights on Hugging Face under the XingChen-AGI organization.

Specifications

	Xing4.0-29B-A4B
Released	September 22, 2026
Vendor	China Telecom AI (TeleAI)
Lineage	Latest generation of TeleChat
Architecture	Mixture-of-experts
Total parameters	29B
Active per token	4B
Native context	256K (extensible to 512K)
License	Apache 2.0
Weights	XingChen-AGI on Hugging Face and GitHub
Published variants	Base, FP8, GGUF
VRAM, unquantized	~62.9 GB
VRAM, low-bit quantized	~15 GB (vendor claim)
Training hardware	Huawei Ascend NPU platform, end to end
Positioning	Agentic execution: planning, tool use, long documents

The single-GPU claim, checked

The headline — “runs on one consumer GPU” — is true with a caveat that matters.

MoE models are memory-heavy and compute-light. You must resident all 29B parameters; you only compute 4B per token. At FP16 that is about 62.9 GB, which is two A100 80GBs’ worth of comfort or an H100. At 4-bit, the rule of thumb of ~0.6 GB per billion parameters lands you near 17-18 GB before KV cache, and China Telecom’s own optimization work claims ~15 GB. That fits:

16 GB cards (RTX 4060 Ti 16GB, RTX 5060 Ti 16GB, A4000) — tight, short context only
24 GB cards (RTX 3090, RTX 4090, RTX 5090 at 32 GB) — comfortable, with room for a real KV cache
Apple Silicon with 32 GB+ unified memory — viable via the GGUF build in llama.cpp or LM Studio

The KV cache is the part people underestimate. A 256K context is the model’s headline feature and also the thing that will exhaust a 16 GB card instantly. Budget context to your VRAM: on 24 GB expect tens of thousands of tokens of working context at 4-bit, not 256K. The 256K and 512K numbers are architectural capability, achievable on server hardware.

Benchmarks and how to read them

China Telecom publishes harness-specific numbers rather than a single score, which is the honest way to do it for an agent model:

Benchmark	Harness	Settings disclosed
SWE-bench Verified	SWE-agent	temp 1.0, top_p 0.95, rep penalty 1.05, 210K context
SWE-bench Multilingual	SWE-agent	same settings, 210K context
Terminal-Bench 2.1	terminus-2	temp 0.8, 64K output cap, 24-hour timeout
Claw-Eval	official harness	temp 0.8, top_p 0.95, max_tokens 16384, 256K context

Third-party coverage reports coding results ahead of Gemma 4 and Qwen3.6 on several of these. Treat that as vendor-adjacent until Artificial Analysis or a comparable independent evaluator publishes — as of September 26, 2026, Xing4.0-29B-A4B has no independent Intelligence Index placement.

The disclosure of harness and sampling settings is itself worth noting. Agent benchmark scores move enormously with harness choice; a model reporting “SWE-bench Verified: X%” without naming the harness is not reporting anything. See why agent scores differ between harness and model.

Where it actually gets used

China Telecom has deployed it into its group-level customer service platform and into interactive home-service environments, for inquiries that need multi-step reasoning and tool calls rather than retrieval-and-answer. That is a useful signal: an operator putting its own open-weight model into a production support stack serving a national telco customer base is a stronger reliability claim than a benchmark table.

For outside builders, the realistic use cases:

Local coding agent on a workstation, where 4B active parameters means interactive latency without an API bill
Document-heavy analysis where the 256K context avoids chunking, on a server with enough VRAM
Air-gapped or data-residency-constrained deployments where Apache 2.0 weights and no vendor call-home are hard requirements
First hop in a router that escalates hard tasks to a frontier model — the standard 2026 cost pattern

The Ascend angle

Trained entirely on Huawei Ascend NPUs, with no Nvidia in the loop. That is the part with strategic weight.

Through 2024-2025 the working assumption was that Ascend was adequate for inference and unproven for frontier-scale training. A 29B MoE agent model with competitive coding scores, shipped with FP8 and GGUF builds and deployed into a telco’s production support platform, is evidence that assumption has expired. Paired with Huawei’s Atlas 960E SuperPod at the cluster tier — see Atlas 960E vs Nvidia Vera Rubin NVL72 — the domestic stack now covers train and serve.

It also explains the release strategy. Apache 2.0, GGUF builds on day one, a GitHub org, English-language model cards: this is a model designed to be adopted outside China, which builds the Ascend ecosystem’s credibility as a side effect.

How it compares to the other small-active MoE models

The single-GPU agent tier got crowded in September 2026. Head-to-head analysis is in Xing4.0-29B-A4B vs Qwen3.6-27B vs gpt-oss-20b, and the broader field is ranked in best open-weight models for a single GPU. Short version: Qwen3.6-27B is the strongest general default, gpt-oss-20b is the lightest at ~12 GB quantized, and Xing4.0-29B-A4B is the one built specifically as an execution agent with the longest native context of the three.

Note that MiMo-V2.6-Flash is not in this tier despite frequent confusion — at 309-310B total parameters it needs datacenter memory even though only ~15B are active. Active parameter count sets speed; total parameter count sets whether it fits.

How to run it

# GGUF via Ollama / llama.cpp — easiest path on a 24 GB card
huggingface-cli download XingChen-AGI/Xing4.0-29B-A4B-GGUF

# FP8 via vLLM on a server GPU
vllm serve XingChen-AGI/Xing4.0-29B-A4B-FP8 --max-model-len 131072

Start context at 32K-64K on consumer hardware and raise it until you hit an OOM, rather than requesting 256K and debugging a crash.

Last verified: September 26, 2026. Benchmark figures are China Telecom’s own published results; no independent Artificial Analysis placement exists yet.