Xing4.0-29B-Intent Locally: 62 GB of Weights, No GGUF, and One Honest Alternative

Verdict first: China Telecom's Xing4.0-29B-Intent is a real, verifiably open model — Apache-2.0, weights on Hugging Face since Oct 15, config audited. But "open" does not mean "runnable on your machine": 62.43 GB of BF16 weights, no official GGUF, and a custom architecture mean nobody is running this on a consumer box today. If you want it locally, you're waiting for community quants or provisioning 80 GB of VRAM. If you want the idea, not the weights, fine-tune your own intent model on its 2.75M-example corpus instead.

After we covered last week's decision-model pile-up, one model stood out for a boring reason: it's the only one with downloadable weights. This page is the practical readout — what the repo actually contains, what hardware it demands, and whether running it yourself makes sense versus just doing the smarter thing and training your own.

What's actually in the repos

I pulled the Hugging Face org XingChen-AGI directly (via mirror; the org is easy to miss because searching "Xing4.0" alone often surfaces nothing — the org name is the findable token). Three assets:

• Xing4.0-29B-Intent — the intent model. 41 safetensors shards totaling 62.43 GB, Apache-2.0, created Oct 15. At publication it showed 759 downloads and 18 likes, which tells you roughly how many people outside China's telecom ecosystem have noticed it: almost none.
• Xing4.0-29B-A4B — the base model it fine-tuned from, a Mixture-of-Experts with 29B total parameters and around 4B active per token. Also 62.43 GB on disk (weights, not activations — MoE saves compute, not disk). Created Sep 16, Apache-2.0, no official GGUF.
• Xing4.0-29B-Intent-oss — the training corpus: 2.75M examples, also open. This is quietly the most valuable item of the three.

The config I pulled confirms the architecture claims from the official release: 40 layers, hidden size 3584, MoE routing 4 experts per token, and — the number that actually matters for deployment — a 262,144-token context window. That's 256K in a 29B model, which is unusual and hints at why this exists: long transcripts of messy human dialogue, not code.

The hardware math nobody in the press release mentions

62.43 GB of weights means BF16 precision (29B parameters × 2 bytes ≈ 58 GB theoretical; add embeddings and the mismatch closes). To load it in one piece you need roughly 64 GB+ of VRAM — an A100 80GB or H100 territory, or a multi-GPU rig. A single RTX 4090 (24 GB) holds about a third of it. MoE's "4B active" makes inference cheaper per token, not smaller in memory — a distinction half the launch coverage in Chinese glosses over, and one English posts haven't corrected yet because, as far as I can find, nobody has written a deployment guide for this model in English at all.

The usual consumer escape hatch is GGUF quantization via llama.cpp. Here's the catch: model_type is xing4_0, a custom architecture, and llama.cpp's conversion tooling only supports architectures someone has bothered to implement. Until a community quant appears (or TeleAI ships official GGUFs, which it hasn't for either the base or the intent model), your realistic local options are: a vLLM or SGLang deployment on an 80 GB card, or CPU offloading that will make a "sub-second latency" intent model embarrassingly slow. The official release mentions prebuilt inference images on ModelScope — that's the vendor-sanctioned path, and it assumes you're comfortable in a China-hosted registry.

So should you run it locally?

Honest ranking of your options:

1. Don't run this specific model — train yours on its corpus. The intent model is a fine-tune of Xing4.0-29B-A4B on 2.75M labeled intent examples, all Apache-2.0. If your use case is ticket routing or tool selection, your schema is different from China Telecom's call-center schema anyway. A 4B–8B dense model fine-tuned on your own data will be smaller, faster, and yours. The corpus is the actual gift here.

2. Rent the VRAM, eval it, decide. An A100 hour is a dollar or two; pulling 62 GB and running the model against your own intent tests costs less than a coffee and answers more than any benchmark press release — a lesson this site keeps repeating, most recently about decision-model pricing and Microsoft's $0.042/M anchor, where the released price told us more than any leaderboard.

3. Wait for quants if you must have it on-prem. Watch for community GGUF conversions; given the model's near-zero English-community traction (759 downloads), I wouldn't hold my breath. Low attention means low quant priority — that's just how the llama.cpp ecosystem allocates effort.

My on-the-record prediction: the 29B-A4B base becomes quietly popular in Chinese enterprise fine-tuning pipelines within months, while the English-speaking world discovers the corpus first, the model second. If someone ships a working GGUF before that, the calculus changes and I'll update this page — the numbers here were pulled from the repos directly, so re-verification is cheap.

Dated caveat: correct as of October 23, 2026. All figures pulled live from Hugging Face org XingChen-AGI via mirror API (repo dates, shard counts, file sizes, config.json) — not from press releases. ModelScope mirrors confirmed at the same date. GGUF availability, downloads, and tooling support will move; this page reflects the repos as they stood today.

Sources: Hugging Face (via hf-mirror) API pulls, Oct 23 — XingChen-AGI/Xing4.0-29B-Intent (62.43 GB / 41 shards, Apache-2.0, config.json: 40 layers, 262,144-token context, MoE 4 experts/token), XingChen-AGI/Xing4.0-29B-A4B (62.43 GB / 12 large shards, created Sep 16), XingChen-AGI/Xing4.0-29B-Intent-oss (2.75M-example corpus); ModelScope repo listings; our decision-model week breakdown for the category context.