File size: 3,961 Bytes
1511173
 
 
 
 
 
 
6b80065
1511173
 
 
 
 
24bc39f
1511173
24bc39f
 
7d471cb
 
24bc39f
 
 
 
 
 
1511173
 
b820a1a
 
1511173
 
7d471cb
2b7cd97
 
 
 
 
 
 
 
 
 
 
1511173
24bc39f
2b69d13
 
b820a1a
2b69d13
b820a1a
9a68d2e
 
 
1511173
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
---
base_model:
- 152334H/miqu-1-70b-sf
- lizpreciatior/lzlv_70b_fp16_hf
language:
- en
library_name: transformers
license: other
quantized_by: mradermacher
tags:
- mergekit
- merge
---
## About

weighted/imatrix quants of https://huggingface.co./wolfram/miquliz-120b-v2.0

<!-- provided-files -->
static quants are available at https://huggingface.co./mradermacher/miquliz-120b-v2.0-GGUF
## Usage

If you are unsure how to use GGUF files, refer to one of [TheBloke's
READMEs](https://huggingface.co./TheBloke/KafkaLM-70B-German-V0.1-GGUF) for
more details, including on how to concatenate multi-part files.

## Provided Quants

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

| Link | Type | Size/GB | Notes |
|:-----|:-----|--------:|:------|
| [GGUF](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-IQ1_S.gguf) | i1-IQ1_S | 25.7 | for the desperate |
| [GGUF](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-IQ2_XXS.gguf) | i1-IQ2_XXS | 32.2 |  |
| [GGUF](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-IQ2_XS.gguf) | i1-IQ2_XS | 35.8 |  |
| [GGUF](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q2_K.gguf) | i1-Q2_K | 44.6 | IQ3_XXS probably better |
| [GGUF](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-IQ3_XXS.gguf) | i1-IQ3_XXS | 47.3 | fast, lower quality |
| [PART 1](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q3_K_XS.gguf.split-aa) [PART 2](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q3_K_XS.gguf.split-ab) | i1-Q3_K_XS | 49.3 |  |
| [PART 1](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q3_K_S.gguf.split-aa) [PART 2](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q3_K_S.gguf.split-ab) | i1-Q3_K_S | 52.2 | IQ3_XS probably better |
| [PART 1](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q3_K_M.gguf.split-aa) [PART 2](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q3_K_M.gguf.split-ab) | i1-Q3_K_M | 58.2 | IQ3_S probably better |
| [PART 1](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q3_K_L.gguf.split-aa) [PART 2](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q3_K_L.gguf.split-ab) | i1-Q3_K_L | 63.4 | IQ3_M probably better |
| [PART 1](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q4_K_S.gguf.split-aa) [PART 2](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q4_K_S.gguf.split-ab) | i1-Q4_K_S | 68.7 | almost as good as Q4_K_M |
| [PART 1](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q4_K_M.gguf.split-aa) [PART 2](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q4_K_M.gguf.split-ab) | i1-Q4_K_M | 72.6 | fast, medium quality |
| [PART 1](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q5_K_M.gguf.part1of2) [PART 2](https://huggingface.co./mradermacher/miquliz-120b-v2.0-i1-GGUF/resolve/main/miquliz-120b-v2.0.i1-Q5_K_M.gguf.part2of2) | i1-Q5_K_M | 85.4 | best weighted quant |


Here is a handy graph by ikawrakow comparing some lower-quality quant
types (lower is better):

![image.png](https://www.nethype.de/huggingface_embed/quantpplgraph.png)

And here are Artefact2's thoughts on the matter:
https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9

<!-- end -->