Local LLM Hardware

Local LLM on Apple M4 Ultra

128GB VRAM · Unified LPDDR5X · 1000 GB/s memory bandwidth · 80W TDP · released 2025

At a glance

VRAM128 GB
MemoryUnified LPDDR5X
Memory bandwidth1000 GB/s
TDP80 W
Released2025
2026 price (used / new)~$4500
TierApple Silicon workstation 2025
Best for128GB unified, 70B-405B at FP16

What models fit on Apple M4 Ultra?

Memory bandwidth is the bottleneck for LLM token generation. The Apple M4 Ultra's 1000 GB/s bandwidth gives roughly the throughput below at Q4_K_M quantization. Numbers are conservative and assume a single request (batch=1), 2048-token context.

ModelQuantizationFits?Approx tokens/sec
Llama 3.2 1BQ4_K_MYes (0.8 GB)~180 tok/s
Llama 3.1 8BQ4_K_MYes (~5 GB)~60
Llama 3.1 8BQ8_0Yes (~8 GB)~45
Qwen 2.5 14BQ4_K_MYes (~9 GB)~35
Mistral Nemo 12BQ4_K_MYes (~8 GB)~40
Qwen 3 32BQ4_K_MYes (~20 GB)~18
DeepSeek R1 distilled 32BQ4_K_MYes (~20 GB)~18
Llama 3.1 70BQ4_K_MYes (~42 GB)~10
Llama 4 Scout 109B (MoE)Q4_K_MYes (~64 GB)~8

Token estimates use a simple bandwidth heuristic (~tokens/sec ≈ bandwidth_GB/s × model_efficiency). Real numbers vary with model architecture, batch size, and KV-cache size. Treat as ballpark.

Build recommendations for Apple M4 Ultra

Best model to download first

Llama 3.1 8B for everyday chat and coding; Qwen 3 32B for top quality on 24GB. Start with ollama pull llama3.1:8b.

Recommended inference backend

For Apple Silicon, use MLX (best performance) or Ollama (simplest setup). Mullama works via llama.cpp but won't beat MLX on Apple Silicon.

Other GPUs to consider

Sources

[1] apple.com/mac-studio