inferencerlabs/DeepSeek-V4-Flash-MTP-DSpark-MLX

DeepSeek-V4-Flash-0731 DSpark

See DeepSeek-V4-Flash with MTP in action: demonstration videos

This draft model contains the extracted DSpark layers to be used alongside the DeepSeek-V4-Flash-0731-MLX model as a speculative decoder for improved performance.

Tested on an M3 Ultra 512 GiB RAM using Inferencer app v2.3.1

Tetris HTML

Text inference	~31 tokens/s @ 500 tokens ~148.5 GiB
Text inference (with MTP)	~35.8 tokens/s @ 500 tokens ~152.9 GiB
Text inference (with DSpark)	~53.1 tokens/s @ 500 tokens ~165.6 GiB

Tell me about Paris

Text inference	~30.5 tokens/s @ 500 tokens ~148.5 GiB
Text inference (with MTP)	~37.3 tokens/s @ 500 tokens ~152.9 GiB
Text inference (with DSpark)	~31.3 tokens/s @ 500 tokens ~165.6 GiB

Enable speculative decoding in Inference Contols:

Extracted with a modified version of MLX-VLM

For more details see our demonstration videos or visit 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.