DeepSeek-V4.1-Flash EXL3 2.0 bpw MCG

EXL3 pack of deepseek-ai/DeepSeek-V4.1-Flash. Routed experts are 2.0 bpw MCG. Other tensors keep the source dtypes. Engram tables stay on NVMe at serve time.

This pack exists so 2× NVIDIA DGX Spark (GB10) can serve the model at tensor-parallel 2. Native MXFP4/MXFP8 is about 511 GB and does not fit 2× Spark UMA.

Download

hf download sfxnz/DeepSeek-V4.1-Flash-EXL3 --revision 2.0bpw-mcg

The two Engram shards are about 95 GiB each. Keep Hugging Face xet enabled. Do not set HF_HUB_DISABLE_XET.

Serve

The 2× DGX Spark cookbook is sfxnz/DeepSeek-V4.1-Flash-EXL3-vLLM-2x-DGX-Spark. Clone that repo, download this revision, build the image on both nodes, and run ./run.sh.

Pack

Field Value
Source deepseek-ai/DeepSeek-V4.1-Flash commit dba1be0a40aa45a94ad051997016db3960a90277
Quant EXL3 2.0 bpw, codebook MCG, quant_method=exl3
Shards model-00001-of-00048.safetensors through model-00048-of-00048.safetensors
Size on disk about 334 GB

Model weights are MIT from DeepSeek. vllm-exl3 is AGPL-3.0. The recipe image clones it at build.

Rebuild

The GitHub recipe documents tools/quantize_experts_exl3.py and tools/assemble_pack.sh if you want to rebuild from the official snapshot.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for sfxnz/DeepSeek-V4.1-Flash-EXL3

Finetuned
(11)
this model