Releasing DeepSeek-V4-Flash-REAP-145B-A13B, an LLM that runs on a single DGX Spark, on Hugging Face

Ludo Technologies has released DeepSeek-V4-Flash-REAP-145B-A13B, a compressed open-weight version of DeepSeek-V4-Flash, on Hugging Face.

The model is derived from DeepSeek-V4-Flash-0731 (284B total parameters) using REAP (Router-weighted Expert Activation Pruning), an expert pruning method. By keeping only the 128 most important experts out of 256 in each layer, we reduced the total parameter count to about 145B and the on-disk size from 167 GB to 82.4 GB. The number of active parameters per token (13B) and the routing behavior are unchanged from the original.

Its key feature is that it runs on a single NVIDIA DGX Spark with 128 GB of unified memory. The full-size model does not fit on that machine, but the compressed model runs on stock vLLM with no patches and decodes at roughly 17 tokens per second.

Calibration used a mix of English chat, code, and Japanese prose and instructions. We found that calibrating on English only caused the model to answer Japanese questions in Chinese, so including Japanese in the calibration data was essential to preserving Japanese output quality after pruning.

The model is released under the MIT license. For downloads and technical details, see Hugging Face (huggingface.co/ludo-tech/DeepSeek-V4-Flash-REAP-145B-A13B) and our blog post.