---
language: en
tags:
- quantized
- mlx
base_model:
- deepseek-ai/DeepSeek-V4-Flash
base_model_relation: quantized
library_name: mlx
pipeline_tag: text-generation
---
# DeepSeek-V4-Flash
See DeepSeek-V4-Flash in action: [demonstration videos](https://youtube.com/xcreate)
   
For improved performance, download and enable a compatible [MTP speculative decoder](https://huggingface.co/models?search=inferencerlabs/deepseek-v4-flash-mtp-mlx).

#### Tested on an M3 Ultra 512 GiB RAM using [Inferencer app v1.11.7](https://inferencer.com)
<table style="border-collapse: collapse; border: none; text-align:left; margin-top:10px; margin-bottom:0px;">
  <tr><td><strong>Text inference</strong></td><td>~31 tokens/s @ 1000 tokens ~148.5 GiB</td></tr>
  <tr><td><strong>Text inference (with MTP)</strong></td><td>~37.3 tokens/s @ 1000 tokens ~152.9 GiB</td></tr>
</table>

<p style="margin-bottom:0px;">
<strong>Q9 typically achieves near lossless accuracy in our coding test</strong>

In this build, the 4-bit pre-quantized weights of the base model were repacked (rather than dequantized and re-quantized to 9-bit), as this approach performed slightly better in our initial coding tests. All remaining weights were quantized to 9-bit.
</p>

![Screenshot](https://cdn-uploads.huggingface.co/production/uploads/688479d616f1ec82fa645019/ueGDAcsebpcWYhuU9Gvn4.jpeg)

##### Quantized with a modified version of [MLX](https://github.com/ml-explore/mlx)
##### For more details see our [demonstration videos](https://youtube.com/xcreate) or visit [DeepSeek-V4-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash).

## 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.