File size: 1,785 Bytes
2b8b8fa
 
 
 
 
 
 
fa76987
2b8b8fa
 
 
 
 
 
 
 
 
 
 
 
9223012
 
3d6a850
fa76987
718faee
9223012
 
 
3d6a850
9223012
3d6a850
f6db853
3d6a850
 
9223012
 
2b8b8fa
 
 
 
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
---
language: en
library_name: mlx
tags:
- mlx
pipeline_tag: text-generation
---
**See DeepSeek-V3.2-Speciale 5.5bit MLX in action - [demonstration video](https://youtu.be/b6RgBIROK5o)**

*q5.5bit quant typically achieves 1.141 perplexity in our testing*
| Quantization | Perplexity |
|:------------:|:----------:|
| **q2.5**     | 41.293     |
| **q3.5**     | 1.900      |
| **q4.5**     | 1.168      |
| **q5.5**     | 1.141      |
| **q6.5**     | 1.128      |
| **q8.5**     | 1.128      |

## Usage Notes
    
#### M3 Ultra 512GB RAM using [Inferencer app v1.7.3](https://inferencer.com)
* Expect ~16.5 tokens/s @ 1000 tokens
* Memory usage: ~450 GB
  * For a larger context window (>11k tokens) you can expand the RAM limit:
    ```bash
    sudo sysctl iogpu.wired_limit_mb=507000
    ```

#### M3 Ultra 512GB RAM connected to MBP 128GB RAM using [Inferencer app v1.7.3](https://inferencer.com) with LAN distributed compute
* Expect ~13.7 tokens/s @ 1000 tokens
* Example memory usage: MBP ~20GB + Mac Studio ~430GB
  * More RAM available for larger context window using this method

##### Quantized with a modified version of [MLX](https://github.com/ml-explore/mlx) 0.28
##### For more details see [demonstration video - coming soon](https://youtu.be/b6RgBIROK5o) or visit [DeepSeek-V3.2-Speciale](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale).

## Disclaimer

We are not the creator, originator, or owner of any model listed. Each model is created and provided by third parties. Models may not always be accurate or contextually appropriate. You are responsible for verifying the information before making important decisions. We are not liable for any damages, losses, or issues arising from its use, including data loss or inaccuracies in AI-generated content.