Apple M5 Ultra

Apple M5 Ultra (512GB) is a professional workstation Apple Silicon chip based on the Apple M5 Ultra architecture. Released in 2026. It features 512 GB of LPDDR5X unified memory at 1200 GB/s memory bandwidth. As an Apple Silicon chip, its memory is unified between CPU and GPU, so the full 512 GB can be allocated to model weights. MLX gives the best performance on Apple Silicon; llama.cpp Metal backend is a solid alternative. Both are well-supported by Ollama.

For local LLM inference, this GPU runs 90 of the 94 models we track natively in VRAM at 8K context. The largest model it handles in VRAM is MiMo V2.5 Pro (16.9 t/s at Q2_K). It can run all tracked models including 405B-class frontier models entirely in VRAM. On Ornith 1.5 397B (MoE), it achieves approximately 19.3 tokens per second at Q8_0 quantization.

Apple's Metal backend is fully supported by MLX and llama.cpp, giving excellent performance on macOS. Among workstation GPUs, it sits above Apple M3 Ultra (512GB) and Apple M5 Ultra (256GB) in performance, but below NVIDIA RTX Pro 6000.

Apple M5 Ultra (256GB) is a professional workstation Apple Silicon chip based on the Apple M5 Ultra architecture. Released in 2026. It features 256 GB of LPDDR5X unified memory at 1200 GB/s memory bandwidth. As an Apple Silicon chip, its memory is unified between CPU and GPU, so the full 256 GB can be allocated to model weights. MLX gives the best performance on Apple Silicon; llama.cpp Metal backend is a solid alternative. Both are well-supported by Ollama.

For local LLM inference, this GPU runs 81 of the 94 models we track natively in VRAM at 8K context. The largest model it handles in VRAM is Nemotron 3 Ultra 550B-A55B (13.6 t/s at Q2_K). It can run all tracked models including 405B-class frontier models entirely in VRAM. On Ornith 1.5 397B (MoE), it achieves approximately 42.3 tokens per second at Q3_K_M quantization.

Apple's Metal backend is fully supported by MLX and llama.cpp, giving excellent performance on macOS. Among workstation GPUs, it sits above Apple M3 Ultra (256GB) and Apple M2 Ultra (192GB) in performance, but below NVIDIA RTX Pro 6000.

Apple M5 Ultra (96GB) is a professional workstation Apple Silicon chip based on the Apple M5 Ultra architecture. Released in 2026. It features 96 GB of LPDDR5X unified memory at 1200 GB/s memory bandwidth. As an Apple Silicon chip, its memory is unified between CPU and GPU, so the full 96 GB can be allocated to model weights. MLX gives the best performance on Apple Silicon; llama.cpp Metal backend is a solid alternative. Both are well-supported by Ollama.

For local LLM inference, this GPU runs 67 of the 94 models we track natively in VRAM at 8K context. The largest model it handles in VRAM is Step 3.7 Flash (62 t/s at Q2_K). It handles the full range of 70B-class dense models and large MoE models entirely in VRAM, with ample room for long context. On Qwen 3.6 27B, it achieves approximately 17.6 tokens per second at BF16 quantization.

Apple's Metal backend is fully supported by MLX and llama.cpp, giving excellent performance on macOS. Among workstation GPUs, it sits above Apple M3 Ultra (96GB) and NVIDIA RTX 6000 Ada in performance, but below NVIDIA RTX Pro 6000.