[!WARNING] Experimental AFM artifact. This repository is intended for a later version of maclocal-api that is still in development and has not yet been released. This is an AFM/DwarfStar deployment artifact, not a general-purpose GGUF release. It may not work with current public releases or other runtimes, and its behavior, quality, performance, memory use, compatibility, or file layout may change. For stable general-purpose MLX models, use mlx-community instead.

DeepSeek V4 Flash 0731 - DwarfStar GGUF for AFM

AFM project website: maclocal.ai

This repository packages the tested DwarfStar-compatible DeepSeek V4 Flash 0731 artifacts used by AFM. It provides one unambiguous main model and its optional DSpark speculative-decoding companion.

The GGUF conversion and quantization were created by Antirez for the DwarfStar engine. These files are mirrored from antirez/deepseek-v4-gguf for a simpler AFM installation workflow. This repository does not claim a new conversion or quantization.

Files

File	Purpose	Approximate size
`DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf`	Main DwarfStar model	156 GB
`DeepSeek-V4-Flash-DSpark-support-0731.gguf`	Optional DSpark companion	6 GB

AFM 0.9.15

Install the Hugging Face CLI:

curl -LsSf https://hf.co/cli/install.sh | bash

Download the main model to a location with at least 170 GB free:

bash

mkdir -p /Volumes/edata2/models/ds4

hf download \
  scouzi1966/DeepSeek-V4-Flash-0731-DwarfStar-GGUF \
  DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf \
  --local-dir /Volumes/edata2/models/ds4

Run AFM:

bash

afm mlx \
  -m /Volumes/edata2/models/ds4/DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf \
  -w

AFM reads the GGUF metadata and automatically selects DwarfStar. The --mlx-runtime dwarfstar option is not required.

Optional DSpark

Download the companion:

bash

hf download \
  scouzi1966/DeepSeek-V4-Flash-0731-DwarfStar-GGUF \
  DeepSeek-V4-Flash-DSpark-support-0731.gguf \
  --local-dir /Volumes/edata2/models/ds4

Enable it explicitly for greedy speculative decoding:

bash

afm mlx \
  -m /Volumes/edata2/models/ds4/DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf \
  --dspark-support /Volumes/edata2/models/ds4/DeepSeek-V4-Flash-DSpark-support-0731.gguf \
  --temperature 0 \
  -w

DSpark performance depends on draft-token acceptance and may be slower for low-acceptance prompts. The default DwarfStar path remains the reference.

Requirements

Apple Silicon Mac
AFM 0.9.15 or newer
Approximately 256 GB unified memory or more is recommended for this MXFP4 variant
At least 170 GB free disk space for the main model and download overhead

Integrity

SHA-256 / Hugging Face LFS OIDs:

text

0e3a161b670f686128ec5f92a601dfde616a37bf5e7e48999fa2d32471b57ec6  DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf
7e319924541db3f7a163ed7e11d7532a70d48228ab59d36cb81e1d4511885360  DeepSeek-V4-Flash-DSpark-support-0731.gguf

Attribution and license

Base model: deepseek-ai/DeepSeek-V4-Flash
DwarfStar conversion: antirez/deepseek-v4-gguf
Runtime: antirez/ds4

The model card of the source DwarfStar repository identifies the artifacts as MIT and notes that the base-model copyright remains with DeepSeek. Users are responsible for complying with the source model's terms.