AI station / mini-PC
Mac Studio (M3 Ultra, 256 GB) — Local LLM Performance
Mac Studio (M3 Ultra, 256 GB) is a Desktop running macOS, with 256 GB unified memory.
Specifications
- Memory for models
- 256 GB unified
- Memory bandwidth
- 819 GB/s
- Memory upgradeable
- No — soldered, decided at purchase
- Chip
- Apple M3 Ultra
- Architecture
- ARM, 3nm TSMC
- Form factor
- Desktop
- Operating system
- macOS
- Released
- 2025
- Llama 3.1 8B Q4_K_M
- 43–90 tok/s
Worth knowing before you buy
- The largest single-box unified memory pool available, but at roughly half the memory bandwidth of a high-end discrete GPU — very large models load, then run slowly.
- Memory is soldered; the configuration chosen at purchase is permanent.
Models on the Mac Studio M3 Ultra 256 GB
- Yi 1.5 FamilyYes — comfortably
- VibeThinkerYes — comfortably
- Aya 3B (Tiny Aya)Yes — comfortably
- StarCoder 2Yes — comfortably
- SmolLM3Yes — comfortably
- SmolLM2Yes — comfortably
- Qwen3-CoderYes — comfortably
- Qwen3.8Yes
- Qwen 3.6Yes — comfortably
- Qwen 3.5Yes — comfortably
- Qwen 3Yes
- Qwen 2.5 VLYes — comfortably
- Qwen 2.5 FamilyYes — comfortably
- Poolside Laguna XS 2.1Yes — comfortably
- Phi-4 MiniYes — comfortably
- Phi-4 FamilyYes — comfortably
- Phi 3.5 FamilyYes — comfortably
- OLMo 2Yes — comfortably
- North Mini CodeYes — comfortably
- Nex-N2Yes — comfortably
- Nex-N2.5Yes — comfortably
- Nemotron Cascade 2Yes — comfortably
- Nemotron 70BYes — comfortably
- Nemotron 3 SuperYes — comfortably
- Nemotron 3 Nano OmniYes — comfortably
- Mistral Small 4Yes — comfortably
- Mistral Small 3.1Yes — comfortably
- Mistral Small 3.2Yes — comfortably
- Mistral FamilyYes — comfortably
- MinistralYes — comfortably
- Ministral 3Yes — comfortably
- MiniMax M3Yes, but it is tight
- MiniMax M2.7Yes
- MiniMax M2.5Yes
- MiniCPM-VYes — comfortably
- MiMo-V2.5Yes, but it is tight
- Magistral SmallYes — comfortably
- Llama 4Yes — comfortably
- Llama 4.5Yes — comfortably
- Llama 3.3Yes — comfortably
- Llama 3.2 VisionYes — comfortably
- Llama 3.2 FamilyYes — comfortably
- Llama 3.1 FamilyYes — comfortably
- LFM2.5Yes — comfortably
- InternLM 3Yes — comfortably
- IBM Granite 4.0Yes — comfortably
- IBM Granite 4.2Yes — comfortably
- IBM Granite 4.1Yes — comfortably
- Granite 3.0Yes — comfortably
- GPT-OSSYes — comfortably
- GLM-4.7 / GLM-Z1Yes — comfortably
- GLM-5 / GLM-5.1No
- GLM-5.3-FlashYes — comfortably
- Gemma 4Yes — comfortably
- Gemma 3nYes — comfortably
- Gemma 3Yes — comfortably
- Gemma 2 FamilyYes — comfortably
- Falcon 3Yes — comfortably
- EXAONE 3.5Yes — comfortably
- DevstralYes — comfortably
- DeepSeek V4Yes
- DeepSeek V4.1No
- DeepSeek R1Yes — comfortably
- DeepSeek-OCRYes — comfortably
- Cosmos 3Yes — comfortably
- Command R FamilyYes — comfortably
- Cogito v1Yes — comfortably
- CodestralYes — comfortably
- Bonsai 27BYes — comfortably
- BitNet b1.58Yes — comfortably
- Aya ExpanseYes — comfortably