# Open-source method (MACaQ) beats Unsloth GGUFs! — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/b7005dab446bd8de

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-11 17:32 ingested 2026-11-12 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xto0fa/opensource_method_macaq_beats_unsloth_ggufs/)

## retained text



```
Some results for those using GGUFs and llama.cpp:

# GGUF quality: MACaQ vs. Unsloth and Bartowski

**Qwen3-4B (IQ4\_XS)**

|Quantization|GGUF size|Mean KL ↓|
|:-|:-|:-|
|**MACaQ**|**2.260 GB**|**0.02804**|
|Bartowski|2.271 GB|0.03443|
|Unsloth|2.271 GB|0.03663|

**23.5% lower KL divergence than Unsloth** and **18.6% lower than Bartowski**, at virtually identical file sizes.

**Qwen3.6-35B-A3B (IQ4\_XS)**

|Quantization|GGUF size|Mean KL ↓|
|:-|:-|:-|
|**MACaQ**|18.86 GB|**0.018696**|
|Unsloth (UD-IQ4\_XS)|17.73 GB|0.020954|
|Bartowski|19.70 GB|0.021927|

**Qwen3.8-27B (IQ4\_XS)**

|Quantization|GGUF size|Mean KL ↓|
|:-|:-|:-|
|**MACaQ**|**14.96 GB**|**0.036784**|
|Bartowski|15.48 GB|0.104170|

\~**65% lower KL divergence than Bartowski**!

# Supported GGUF formats (llama.cpp)

|Weight bits|Formats|
|:-|:-|
|2-bit|`IQ2_XXS`, `IQ2_XS`|
|3-bit|`IQ3_XXS`, `IQ3_S`|
|4-bit|`IQ4_NL`, `IQ4_XS`|

Code, llama.cpp fork, and benchmarks: [https://github.com/huawei-csl/MACaQ](https://github.com/huawei-csl/MACaQ)

Cool to see Huawei working on such stuff and sharing some of their internal tools that are helpful for us. Maybe they can build an affordable GPU for us local LLM folks sometime 😂

They also report quantizing a **122B-parameter model in 64 minutes using just 32.5 GB of GPU memory**, which is interesting for a model of that size. I did not try that.

Edit: there is a **paper** on it [https://arxiv.org/abs/2611.10649](https://arxiv.org/abs/2611.10649)

Edit 2: they don't share ready-made gguf files, so I am sharing some myself:  
\- [http://31.165.97.47:20080/macaq-gguf/qwen3-4b-macaq-iq4\_xs-had64-vq2-calsmoke.gguf](http://31.165.97.47:20080/macaq-gguf/qwen3-4b-macaq-iq4_xs-had64-vq2-calsmoke.gguf)

Edit 3: they have this blog post [https://huawei-csl.com/blog/macaq/](https://huawei-csl.com/blog/macaq/) with multi-modal examples, too. The RTN audio comparisons are just too much😂😂

*Processing img 37wuumlwy01i1...*

*Processing img 9s4hxe15z01i1...*

  
The multi-modal stuff is generally insane. Normally 4b already completely goes to s\*\*t, but here it looks almost the same.
```



document fetched 2026-11-12 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xto0fa/opensource_method_macaq_beats_unsloth_ggufs
