A fast fixed schedule with a deliberately narrow contract.

DeepSeek V4 names the model architecture. DwarfStar names AFM's in-process fixed-schedule Metal executor for a specially packed DeepSeek V4 checkpoint. DSpark is an optional speculative support model layered on that executor. They are related, but they are not interchangeable.

The three layers

Runtime selection

--mlx-runtime auto is the default. It selects DwarfStar when the model argument is a local GGUF whose metadata declares the compatible deepseek4 architecture. A remote Hugging Face ID must first be resolved to a local file. Explicit dwarfstar requires a compatible local GGUF and fails early on a directory, missing file, or incompatible architecture; explicit mlx keeps the general MLX path.

# Auto-select DwarfStar for a compatible local GGUF
afm mlx -m /models/deepseek-v4-dwarfstar.gguf

# Require DwarfStar and fail early if the GGUF is incompatible
afm mlx -m /models/deepseek-v4-dwarfstar.gguf   --mlx-runtime dwarfstar

# Force the general MLX runtime for an ordinary MLX checkpoint
afm mlx -m /models/deepseek-v4-mlx   --mlx-runtime mlx

Executor package format

The converter and lower-level AFMKit integration can also describe an AFM-native safetensor executor package. This is packaging reference—not the AFM 0.9.15 CLI launch path above. Its config.json declares all of the following:

The package also needs model.safetensors.index.json, every indexed shard, all real tensor offsets aligned to 32 bytes, and the bundled dwarfstar-template.gguf. The catalog is validated before the engine opens; a nearly compatible package is not accepted.

Build an executor package

afm mlx-convert   --source /path/to/official-DeepSeek-V4-Flash-0731   --output /path/to/deepseek-v4-afm-executor   --profile dwarfstar-executor   --template-gguf /path/to/reference-dwarfstar.gguf

The reference GGUF supplies compact metadata and tensor schema for the bundled template; the converted AFM safetensor shards remain the model payload. Conversion is persistent, shard-streaming, and resumable. Use --overwrite only when you intend to recreate the output directory.

# Upgrade an older AFM executor package in place to the aligned layout
afm mlx-align-executor   --checkpoint /path/to/deepseek-v4-afm-executor

What the runtime does support

• OpenAI-compatible streaming and non-streaming text chat.
• Greedy decoding and temperature sampling with top_p, top_k, min_p, and seed.
• Maximum-token limits, stop sequences, cancellation, token usage, and timing metadata.
• DeepSeek reasoning policy through reasoning_effort, template kwargs, and --no-thinking.
• Resident concurrent sessions with --concurrent N.
• Exact-prefix reuse with --enable-prefix-caching.
• The built-in WebUI, restricted to text input for this runtime.

Limitations matrix

Concurrency and prefix reuse

--concurrent N creates N resident DwarfStar sessions and batches active decode items through the fixed engine. This is ordinary multi-session serving, not the MLX batch API. Each session consumes cache and runtime resources, so increase it only after measuring unified-memory headroom.

Prefix caching is intentionally conservative. AFM reuses a resident session only when the session's complete live token history is a prefix of the new prompt. A shorter common prefix is reported as a miss because DwarfStar would still need to rebuild compressor and KV frontiers.

DSpark speculative decoding

afm mlx -m /path/to/deepseek-v4-dwarfstar.gguf   --mlx-runtime dwarfstar   --dspark-support /path/to/support.gguf   --dspark-draft-tokens 5   --dspark-confidence 0.7

Performance boundary

The repository's M3 Ultra validation found that the fixed executor's roughly two command buffers per generated token removed the graph-fragmentation boundary seen in the native MLX DeepSeek V4 path. The aligned AFM executor measured within the canonical DwarfStar run variance and preserved exact greedy output in the validated bracket.

Those 35–37 token/s measurements are evidence for one DeepSeek V4 Flash 0731 package, prompt, build, and M3 Ultra—not a universal speed guarantee. Checkpoint conversion changes weight ABI and loadability; semantic quality and performance still require validation on the package and Mac you will deploy.

For the complete benchmark record, controlled variables, rejected hypotheses, hashes, and capture inventory, read the DeepSeek V4 performance record.

Choose DwarfStar or MLX