Short answer: yes—the official Muse Spark Open launch article reports two tested alternate-quant runs on a 512GB M3 Ultra Mac. A llama.cpp Q2_K build used about 304GB and ran at 18 tokens/second. An MLX oQ3e build used about 344GB and measured approximately 225 tokens/second prefill and 21 tokens/second generation with a 1,024-token prompt and 128 generated tokens. These are not Meta’s official NVFP4 checkpoint or its 700GB GPU-memory requirement.
Do not mix the artifacts
Official NVFP4: 533.011GB of listed safetensors and at least 700GB VRAM, targeting NVIDIA Blackwell. Tested Mac routes: separate 2-bit GGUF and oQ3e MLX community artifacts that trade precision, format and runtime for a smaller memory footprint.
What the published Mac tests actually show
| Route | Published memory use | Published speed | Test boundary |
|---|---|---|---|
| llama.cpp Q2_K GGUF | About 304GB | 18 tokens/second | M3 Ultra Mac with 512GB unified memory; other prompt details not stated in the article |
| MLX oQ3e | About 344GB | ~225 tok/s prefill; ~21 tok/s generation | M3 Ultra Mac with 512GB; 1,024-token prompt and 128 generated tokens |
| Official Meta NVFP4 | 533.011GB listed files | No Mac result | At least 700GB GPU memory; Blackwell-targeted |
The article says the Q2_K route leaves “plenty of room” for KV cache and gives the 304GB model-use figure. Nominal subtraction leaves 208GB from a 512GB configuration. The oQ3e route leaves 168GB by the same arithmetic. Neither remainder is fully available to the model: macOS, applications, runtime allocations, temporary buffers and memory pressure share the same unified pool.
Unified memory is not dedicated VRAM
Apple unified memory is accessible to CPU and GPU workloads through one pool. That makes very large local model artifacts possible without a bank of discrete accelerators, but it does not mean all 512GB behaves like reserved GPU memory. The operating system and every active process compete for the same capacity.
Apple’s M3 Ultra announcement states that the Mac Studio generation could be configured with up to 512GB unified memory. The Muse Spark launch article’s measurements apply to the tested M3 Ultra configuration—not automatically to a different Mac generation, memory size, runtime build or quant revision.
Why the speed numbers are bounded
- The Q2_K article result does not state prompt length, output length, warmup or repetition count.
- The MLX result does state a 1,024-token prompt and 128 generated tokens, making its prefill and generation figures more reproducible.
- Neither result establishes one-million-token operation, maximum concurrency, sustained thermals or wall power.
- Q2_K and oQ3e are different quantization schemes and runtimes; their numbers are not a direct apples-to-apples performance ranking.
- A newer Mac or newer runtime may perform differently and needs its own measured record.
What smaller Mac memory tiers mean
| Unified memory | 304GB Q2_K footprint | 344GB oQ3e footprint | Evidence-bounded conclusion |
|---|---|---|---|
| 256GB | 48GB short | 88GB short | Neither published footprint fits |
| 384GB | 80GB nominal remainder | 40GB nominal remainder | Arithmetic only; the article did not validate this tier |
| 512GB | 208GB nominal remainder | 168GB nominal remainder | Both routes were tested on this tier |
Memory-fit arithmetic is not a buying recommendation. In particular, the 40GB nominal remainder at 384GB for oQ3e must also hold macOS, runtime state, KV cache and workload buffers, and no reviewed source here validates that configuration.
Reproduce the Mac result
- Record the exact Mac model, chip, GPU-core count and installed unified memory.
- Pin the GGUF or MLX repository revision and exact quant filename.
- Record llama.cpp or MLX/oMLX version plus launch flags.
- State modality, prompt tokens, generated tokens, context setting and reasoning effort.
- Record model memory, peak process memory and system memory pressure.
- Separate prefill from generation throughput.
- Include warmup, repetitions and sampling settings.
- Check output correctness, parser behavior, errors and retries.
- Report sustained thermals and wall power separately if measured.
Decision
A 512GB M3 Ultra Mac has published evidence for running alternate 2-bit Muse Spark Open quants locally. It does not satisfy or replace Meta’s official NVFP4 700GB requirement; it follows a different artifact and runtime path. Treat the Q2_K and oQ3e results as bounded reference measurements, not promises for every prompt, context length or Mac generation.
Muse Spark on Mac Studio: quick answers
Can a 384GB Mac Studio run Muse Spark Open?
The reviewed launch evidence does not validate a 384GB Mac. The reported Q2_K and oQ3e model footprints would leave about 80GB and 40GB of nominal unified memory respectively, but macOS, the runtime, KV cache and workload buffers must share that remainder. Treat 384GB as unverified arithmetic, not a supported recommendation.
Does a 512GB Mac Studio meet Muse Spark’s official NVFP4 requirement?
No. Meta lists 533.011GB of NVFP4 checkpoint files and says that official path needs at least 700GB of GPU memory on Blackwell hardware. The published 512GB M3 Ultra results use separate Q2_K GGUF and oQ3e MLX artifacts.
Which Muse Spark quantizations were tested on the 512GB M3 Ultra?
The launch article reports a llama.cpp Q2_K GGUF path using about 304GB and an MLX oQ3e path using about 344GB. These are different artifacts and runtimes, so their speed figures should not be treated as a direct benchmark comparison.
Will a newer 512GB Mac Studio reproduce the M3 Ultra result?
Not automatically. A newer chip may be a plausible candidate, but the published measurements are for the 512GB M3 Ultra. Reproduction still requires the exact artifact revision, runtime version, flags, prompt and context settings to be recorded and tested.
Choosing the software path? The Muse Spark Open deployment guide separates Meta’s distributed NVFP4 MLX reference from single-Mac llama.cpp, Unsloth and alternate-quant experiments.
Continue the research
- Compare the official NVFP4 files and 700GB requirement
- Understand Mac Studio unified memory for local AI
- Review the current 512GB Mac Studio capacity guide
- Compare the published NVIDIA deployment paths
- Stage a reproducible Muse Spark benchmark record
- Follow qualifying high-memory hardware
Primary sources
- Official launch article with the M3 Ultra llama.cpp and MLX results
- Official Meta NVFP4 repository
- Apple M3 Ultra announcement and 512GB unified-memory specification
Checked October 25, 2026. Unified memory is shared system memory, not dedicated VRAM. Published speed figures apply only to the named artifact, runtime and disclosed test conditions.