Upload complete model
Browse files
README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
library_name: mlx
|
| 4 |
+
base_model: deepseek-ai/DeepSeek-V3.1
|
| 5 |
+
tags:
|
| 6 |
+
- mlx
|
| 7 |
+
pipeline_tag: text-generation
|
| 8 |
+
---
|
| 9 |
+
**See DeepSeek-V3.1 5.5bit MLX in action - [demonstration video](https://youtube.com/xcreate)**
|
| 10 |
+
|
| 11 |
+
*q5.5bit quant typically achieves 1.141 perplexity in our testing
|
| 12 |
+
| Quantization | Perplexity |
|
| 13 |
+
|:------------:|:----------:|
|
| 14 |
+
| **q2.5** | 41.293 |
|
| 15 |
+
| **q3.5** | 1.900 |
|
| 16 |
+
| **q4.5** | 1.168 |
|
| 17 |
+
| **q5.5** | 1.141 |
|
| 18 |
+
| **q6.5** | 1.128 |
|
| 19 |
+
| **q8.5** | 1.128 |
|
| 20 |
+
|
| 21 |
+
## Usage Notes
|
| 22 |
+
|
| 23 |
+
* Built with a modified version of [MLX](https://github.com/ml-explore/mlx) 0.26
|
| 24 |
+
* Memory usage: ~480 GB
|
| 25 |
+
* Expect ~15 tokens/s
|
| 26 |
+
* For more details see [demonstration video](https://youtube.com/xcreate) or visit [DeepSeek-V3.1](https://huggingface.co/deepseek-ai/DeepSeek-V3.1).
|