Comparing 384GB and 512GB for Muse Spark Open? See the measured 512GB M3 Ultra path and the unverified 384GB boundary, including the published 304GB Q2_K and 344GB oQ3e tests and why they do not replace the official 700GB NVFP4 requirement.
Short answer: Mac Studio with M5 Ultra can be configured with 96GB, 256GB or 512GB of unified memory. For large local AI, treat that as a shared system pool—not dedicated VRAM. The 256GB configuration is a practical capacity class for Qwen3.5-122B FP8/Q8 and a tight file boundary for Qwen3.5-397B 4-bit; 512GB is the practical unified-memory class for Qwen3.5-397B FP8/Q8. The official 806.796GB 397B BF16 checkpoint cannot fit in 512GB.
512GB unified memory is not 512GB dedicated GPU VRAM. macOS, applications, CPU work and GPU work share the pool. Confirm the runtime’s usable allocation, model-format support and measured peak before buying.
M5 Ultra specifications that matter for local AI
| Official specification | Mac Studio M5 Ultra | Planning meaning |
|---|---|---|
| Unified memory | 96GB; configurable to 256GB or 512GB | Shared by CPU, GPU, operating system and applications |
| Memory bandwidth | 1.2TB/s | Published SoC bandwidth, not a universal token-rate claim |
| CPU | 30-core; configurable to 36-core | Host preprocessing and runtime work share the same system |
| GPU | 64-core; configurable to 80-core | Core count does not establish model/runtime throughput |
| Neural Engine | 32-core | Availability does not mean every local-LLM runtime uses it |
| Storage | 1TB; configurable up to 16TB | Large artifacts and caches require substantial local storage; SSD is not memory |
Qwen3.5-122B-A10B on 96GB, 256GB or 512GB
| Artifact | Complete files | 96GB unified | 256GB unified | 512GB unified |
|---|---|---|---|---|
| Official GPTQ-Int4 | 78.864GB | File fits, but shared headroom is limited | Comfortable capacity class | Substantial capacity headroom |
| Community Q4_K_M + F16 projector | 77.446GB | File fits, but runtime/vision state shares the pool | Comfortable capacity class | Substantial capacity headroom |
| Official FP8 | 127.163GB | Does not fit | Practical shared-memory capacity | Substantial capacity headroom |
| Community Q8_0 + F16 projector | 130.781GB | Does not fit | Practical shared-memory capacity | Substantial capacity headroom |
| Official BF16 | 250.173GB | Does not fit | File boundary; insufficient shared headroom | Practical shared-memory capacity |
Qwen3.5-397B-A17B on 256GB or 512GB
| Artifact | Complete files | 256GB unified | 512GB unified |
|---|---|---|---|
| Official GPTQ-Int4 | 235.708GB | File boundary; little shared headroom | Practical shared-memory capacity |
| Community Q4_K_M + F16 projector | 245.012GB | Very tight file boundary | Practical shared-memory capacity |
| Official FP8 | 406.152GB | Does not fit | Practical shared-memory capacity, subject to runtime support |
| Community Q8_0 + F16 projector | 422.426GB | Does not fit | Practical shared-memory capacity, subject to runtime support |
| Official BF16 | 806.796GB | Does not fit | Does not fit |
MLX and runtime safeguards
Apple’s MLX documentation explains that arrays live in unified memory and can be operated on by CPU or GPU without the conventional device-copy model. MLX-LM provides an LLM-focused path on Apple silicon, but framework support is not the same as exact Qwen3.5 model, quantization, hybrid-attention, MoE, MTP or multimodal support.
- Choose the exact model repository, revision and format supported by the runtime.
- Verify Qwen3.5 architecture and projector/image/video support at the exact MLX-LM or alternative runtime version.
- Close unrelated memory-heavy applications and record actual available memory before load.
- Fix context allocation, visual workload, output length, batch size and concurrency.
- Measure peak unified memory, swap activity, time to first token, prompt processing, generation rate and errors.
- Do not compare a shared-memory result with dedicated-GPU results without labeling the memory and timing differences.
M5 Ultra versus multi-accelerator systems
A 512GB M5 Ultra system offers one coherent shared-memory machine rather than several discrete accelerator pools, which can simplify capacity planning for supported runtimes. Dedicated HBM/GDDR systems provide different software, memory, interconnect and deployment characteristics. Neither architecture wins universally: compare the exact workload, supported artifact, measured speed, peak memory, power at the wall and complete acquisition/hosting cost.
Continue with the Qwen3.5-122B guide, Qwen3.5-397B guide, 96GB-to-1TB hardware comparison, MI300X guide, workload test, benchmark submission, private watchlist/RSS page or all guides and tools.
Apple and MLX sources
- Apple Mac Studio M5 Max/M5 Ultra technical specifications
- Apple Mac Studio product page
- Apple MLX repository
- MLX unified-memory documentation
- MLX-LM repository
- HighVRAM methodology
Reviewed October 25, 2026. Hardware specifications are Apple-published; MLX behavior is documented by the Apple ML research project. HighVRAM has not independently benchmarked every model, runtime, format, context or multimodal workload. File fit does not establish compatibility, speed, quality, reliability or value.
Cross-platform path: compare the Mac Studio M5 Ultra and RTX 5090 to see which exact Qwen artifacts cross 32GB, 96GB, 256GB and 512GB capacity boundaries—and where shared-memory caveats matter.
Comparing large unified memory for gpt-oss-120b? Use the dedicated hardware guide to keep the Apple runtime path separate from OpenAI’s cited CUDA/H100 path.
Comparing the two workstation paths? Read Mac Studio M5 Ultra vs RTX PRO 6000 Blackwell for the 512GB shared-memory versus 96GB dedicated-CUDA decision.
Compact-system alternative: compare M5 Ultra with DGX Spark across 96GB–512GB versus 128GB, exact model-file boundaries and software support.
Muse Glimmer is a much smaller consumer-class model. Read the Muse Glimmer Mac memory guide for 24GB through 64GB quantized paths; do not transfer Muse Spark Open’s artifacts, requirements or results.