Spaces:

---

kingjones777

/

quant-lab

Like 2

Running

Hi Xing Team,

Thank you so much for reaching out. I genuinely appreciate it, and it means a lot that you took the time to look at my Xing4.0 quant and the recipe and hashes I published with it.

To answer your questions directly: I really enjoy working with the Xing models, and I’m still learning and expanding my knowledge around quantization, optimization, and getting the best possible performance from these models. Any technical information, documentation, recommendations, early guidance, or other resources your team is willing to share would be greatly appreciated. I’m always interested in learning more and improving the quality of the builds I publish.

I would also absolutely be interested in staying in touch.

Please feel free to reach out anytime if your team has something you would like tested, benchmarked, quantized, or evaluated on different hardware. I genuinely enjoy experimenting with new models and pushing them across different platforms and configurations.

Thanks again for reaching out. I’m looking forward to staying connected with the Xing team and continuing to work with your models.

Best,

Myron

Hi Myron,

You mentioned that if the team ever had something to be tested, benchmarked, or quantized, we should reach out — so here it is.

Last Friday we open-sourced Xing4.0-29B-Intent, fine-tuned from the same Xing4.0-29B-A4B base your quants are built on. Same MoE architecture (29B total, ~4B activated per token), so your recipes and calibration workflow should carry over directly. The model is schema-guided: text or conversations in, structured JSON out (category + args), unifying intent recognition with slot filling, NER, relation extraction, and event extraction in one format.

It ranked first on our internal benchmark of 18 general-purpose models (Overall 58 vs. 49 for Gemini-3.6-Flash; 6,000 samples from 48 public datasets — our own test set, so the usual caveats). We also released the full training dataset — 2.75M samples rebuilt from 61 public datasets, 19 domains, 52 languages, Apache-2.0 with upstream licenses preserved — plus a Python SDK.

Two things that may be relevant to you:

• Official FP8 is already up; the official GGUF is still being prepared. If you want to put a ROCm build out before we do — as before, you're the one we'd most like to see it running on AMD.
• If the released dataset fits your calibration work, it's free to use under Apache-2.0.

If anything misbehaves on AMD — won't load, silent errors, odd performance — post it right here. We owed you a faster answer last time; you won't have to wait again.

Best,
Xing teams
China Telecom AI

• Model: https://huggingface.co/XingChen-AGI/Xing4.0-29B-Intent
• FP8: https://huggingface.co/XingChen-AGI/Xing4.0-29B-Intent-FP8
• Dataset: https://huggingface.co/XingChen-AGI/Xing4.0-29B-Intent-oss
• GitHub (incl. SDK): https://github.com/XingChen-AGI/Xing4.0-29B-Intent

On it

done