Muse Spark Open has published local Mac results—but they use alternate 2-bit artifacts, not Meta’s official NVFP4 checkpoint. The official launch article reports a llama.cpp Q2_K run using about 304GB at 18 tokens/second and an MLX oQ3e run using about 344GB at approximately 225 tokens/second prefill and 21 tokens/second generation, both on a 512GB M3 Ultra Mac.
The two published Mac results
| Route | Memory use | Published speed | Test boundary |
|---|---|---|---|
| llama.cpp Q2_K GGUF | About 304GB | 18 tokens/second | 512GB M3 Ultra; prompt and repetition details not stated |
| MLX oQ3e | About 344GB | ~225 tok/s prefill; ~21 tok/s generation | 512GB M3 Ultra; 1,024-token prompt, 128 generated tokens |
| Official Meta NVFP4 | 533.011GB listed files | No Mac result | At least 700GB GPU memory; Blackwell-targeted |
Why “runs on a Mac” needs an artifact label
The Q2_K GGUF and oQ3e MLX builds are separate community quantization and runtime paths. Their smaller footprints do not revise Meta’s 700GB requirement for the official NVFP4 checkpoint. The result is still important: it establishes that a 512GB Apple-silicon system can run useful compressed variants locally.
The nominal headroom is shared
Subtracting the reported footprints from 512GB leaves 208GB for Q2_K and 168GB for oQ3e. That is not dedicated KV-cache space. macOS, applications, model runtime allocations, temporary buffers and memory pressure share the same unified-memory pool.
Apple’s M3 Ultra announcement confirms that its Mac Studio generation could be configured with up to 512GB unified memory. The measurements remain specific to the tested M3 Ultra, artifact and runtime; they do not automatically transfer to another Mac generation or quant revision.
What the measurements do not establish
- The Q2_K result does not disclose prompt length, generated length, warmup or repetitions.
- Neither result proves full one-million-token operation or a maximum safe context.
- Neither result establishes concurrency, sustained thermals or wall power.
- The two runtimes and quant formats are different, so their speeds are not a direct ranking.
- 256GB is smaller than either published model footprint; 384GB fits them arithmetically but was not validated by the cited test.
Use the complete Mac analysis
The complete 512GB Mac Studio guide separates both alternate quants from NVFP4, compares 256GB, 384GB and 512GB tiers, and supplies a nine-step reproduction checklist.
- Check the official NVFP4 files and 700GB requirement
- Choose the matching SGLang, vLLM, MLX, llama.cpp or Unsloth runtime path
- Understand Mac unified memory for local AI
- Review the current 512GB Mac capacity guide
- Stage a reproducible benchmark record
- Follow qualifying high-memory hardware
- Subscribe to the main RSS feed
- Read the latest Weekly Digest synthesis
Sources
- Official launch article with the M3 Ultra llama.cpp and MLX results
- Official Meta NVFP4 repository
- Apple M3 Ultra and 512GB unified-memory announcement
Checked October 25, 2026. Unified memory is shared system memory, not dedicated VRAM. Speed figures apply only to the named artifacts and disclosed test conditions.