One large download. One local command. A temporary path until the next AFM release.

This is the canonical, shareable quickstart for running DeepSeek V4 Flash through AFM's DwarfStar backend today. AFM 0.9.15 expects a local compatible GGUF, so the command asks Hugging Face to resolve the model file first and passes the resulting local path to AFM.

Before you start

• An Apple Silicon Mac with enough unified memory for this DeepSeek V4 configuration.
• At least 156 GB of free storage for the GGUF, plus working headroom.
• AFM 0.9.15 and the current hf command-line client.
• A fast, reliable connection. The first run downloads the model; later runs reuse the cache.

1. Install AFM and the Hugging Face CLI

# AFM — Homebrew
brew install scouzi1966/afm/afm

# Hugging Face CLI
curl -LsSf https://hf.co/cli/install.sh | bash -s

PyPI users can install AFM with pip install macafm==0.9.15. Open a new terminal after installing hf if your shell cannot find the command.

2. Download and run

This version stores the model in the normal Hugging Face cache. Copy the complete block:

afm mlx -w -m "$(hf download \
  scouzi1966/DeepSeek-V4-Flash-0731-DwarfStar-GGUF \
  DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf \
  --cache-dir "$HOME/.cache/huggingface" \
  --quiet)"

AFM recognizes the GGUF's deepseek4 architecture and selects DwarfStar automatically. The -w flag starts the WebUI; open http://127.0.0.1:9999 after startup completes.

3. Put the model on an external disk

Use this variant when your internal disk does not have enough room. Replace /Volumes/edata2 with the mounted path of your disk.

mkdir -p /Volumes/edata2/models/huggingface-cache

afm mlx -w -m "$(hf download \
  scouzi1966/DeepSeek-V4-Flash-0731-DwarfStar-GGUF \
  DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf \
  --cache-dir /Volumes/edata2/models/huggingface-cache \
  --quiet)"

Verify the server

In another terminal, check the health endpoint and the active model:

curl -s http://127.0.0.1:9999/health
curl -s http://127.0.0.1:9999/v1/models

Know the boundary

• DwarfStar is a text-generation backend. Images, PDFs, audio, --vlm, and --media are rejected.
• This path does not provide tool calls, structured-output grammar, or token log probabilities.
• MLX-specific KV controls, MTP, EAGLE3, and MLX batch endpoints do not apply.
• Not every GGUF is compatible. This command names the validated DeepSeek V4 DwarfStar file explicitly.

Where to go next

For runtime selection, concurrency, prefix reuse, DSpark, sampling controls, and the full limitations matrix, read DwarfStar + DeepSeek V4. Release changes are tracked in What's new.