Can a Mac Run Muse Glimmer 30B? 24GB, 32GB, 48GB and 64GB

Yes, an Apple-silicon Mac can run quantized Muse Glimmer 30B, but the right memory tier depends on the runtime. Do not translate Meta’s general 24GB Q4_K_M hardware target into a universal 24GB Mac recommendation. Apple unified memory is shared by the CPU, GPU and operating system; Meta’s SGLang MLX path lists Apple silicon with 48GB or more, and Meta’s LM Studio guidance reports about 24.1GiB for Q4_K_M at one request and 16K context.

Muse Glimmer Mac memory tiers

Mac unified memory	Evidence-bounded role	What must be verified
24GB	Constrained experiment only	Lean runtime, short context, no competing applications, peak memory and stability
32GB	Q4_K_M text candidate	Exact artifact/tag, actual context, runtime overhead and OS pressure
48GB	First SGLang-listed Apple tier	MLX artifact, text-only limitation, cache settings and measured throughput
64GB	Quantized path with more operating room	Whether extra memory solves context or workload needs; BF16 is still a separate larger target
96GB–128GB+	Additional quantized-context and application headroom	Runtime support and measured value; unused capacity is not automatic performance

Why 24GB GPU guidance is not 24GB Mac guidance

Apple silicon uses a unified-memory architecture: the CPU and GPU share the same physical memory. That removes a discrete transfer boundary, but it also means the model does not receive the entire advertised capacity as dedicated GPU memory. macOS, the runtime, KV cache, prompts, applications and tool processes all draw from the shared pool.

Meta’s smaller official text checkpoint is about 17GB and its rough working envelope is approximately 17GB for text, 19GB with vision and 20GB with vision plus the DFlash drafter. Those figures justify a general 24GB hardware target. They do not guarantee that every Mac application stack remains below 24GB. Meta’s LM Studio troubleshooting section reports roughly 24.1GiB for Q4_K_M with one request and a 16K context, already beyond a nominal 24GB pool before treating the rest of the machine as free.

Best-supported Apple paths

Ollama with MLX

Meta documents Ollama 0.32.7 or later and provides a specific Apple-silicon MLX tag:

ollama --version
ollama run muse-glimmer:30b-mlx

This is the shortest local path. Record the resolved tag and context before comparing it with another runtime; the friendly model name is not enough to establish artifact equivalence.

SGLang MLX at 48GB or more

Meta’s deployment table lists three text-only MLX repacks for Apple silicon and identifies 48GB or more of unified memory. The default Q4_K_M-coded path is:

SGLANG_USE_MLX=1 SGLANG_MLX_CACHE_LIMIT_GB=8 \
sglang serve \
  --model-path RadixArk/Muse-Glimmer-q4km-gs128-MLX \
  --served-model-name muse-glimmer \
  --trust-remote-code \
  --reasoning-parser muse \
  --tool-call-parser muse \
  --disable-radix-cache \
  --mem-fraction-static 0.85 \
  --host 0.0.0.0 --port 30000

The MLX backend is text only and does not support speculative decoding in Meta’s documented SGLang path. The cache-limit and disabled-radix-cache settings are requirements, not optional tuning suggestions.

llama.cpp with Metal

Meta lists llama.cpp for CPU, Metal, CUDA, ROCm and Vulkan. Use build b10353 or newer and begin with a controlled 32K single-turn smoke test rather than forcing the 131K ceiling:

llama cli -hf meta-models/Muse-Glimmer-30B-GGUF \
  --hf-file Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf \
  -ngl 99 -c 32768 -st

If a constrained Mac relies on CPU placement or memory pressure, report that behavior explicitly. A run that launches is not the same as a stable, fully resident configuration.

ExecuTorch on Apple silicon

Meta also publishes ahead-of-time ExecuTorch paths for Apple silicon, including prebuilt .pte exports and an MLX build preset. This is a specialized deployment path; it should not be conflated with loading the public GGUF directly in llama.cpp or an Ollama library build.

What about vision and DFlash?

The official llama.cpp GGUF route downloads the separate vision projector with the text model and can use the optional DFlash draft. The SGLang MLX route Meta lists for Apple silicon is text only and has no speculative decoding. Therefore, “runs on a Mac” is incomplete unless the result identifies the runtime, modality and drafter state.

For a 32GB Mac, prove text-only operation first. For vision, DFlash or long context, 48GB or 64GB provides a more defensible starting point, but the exact peak still has to be measured.

A reproducible Mac test

1. Record the Mac model, chip and total unified memory.
2. Close or document competing memory-heavy applications.
3. Pin the runtime version and exact model artifact or resolved library tag.
4. Start with Q4_K_M text-only and a 32K-or-lower context.
5. Run the same prompt at least three times and check the answer.
6. Record memory pressure, swap, peak process memory, time to first token and output rate.
7. Add vision, DFlash or a longer context one variable at a time.
8. Do not compare the result with a discrete GPU unless the artifact and workload match.

Continue planning

• Review the complete 24GB, 32GB and BF16 memory evidence
• Follow the full llama.cpp, Ollama and LM Studio setup guide
• Choose a broader Mac unified-memory tier
• Stage a reproducible Muse Glimmer Mac benchmark
• Inspect screened Mac and high-VRAM hardware observations — confirm configuration, condition, shipping, availability and seller details.
• Stage a private Mac hardware watchlist

Primary sources

• Meta Muse Glimmer deployment documentation
• Meta Muse Glimmer 30B GGUF repository
• Apple: Get started with MLX for Apple silicon

Checked October 25, 2026. Published memory figures and runtime support come from the linked first-party sources. Recommendations above distinguish documented support from candidate configurations and do not claim an unmeasured Mac performance result.