Local LLM Hardware
Local LLM on Apple M3 Ultra
192GB VRAM · Unified LPDDR5 · 800 GB/s memory bandwidth · 60W TDP · released 2024
At a glance
| VRAM | 192 GB |
|---|---|
| Memory | Unified LPDDR5 |
| Memory bandwidth | 800 GB/s |
| TDP | 60 W |
| Released | 2024 |
| 2026 price (used / new) | ~$4000 |
| Tier | Apple Silicon workstation |
| Best for | 192GB unified memory, runs 70B-405B at FP16 |
What models fit on Apple M3 Ultra?
Memory bandwidth is the bottleneck for LLM token generation. The Apple M3 Ultra's 800 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.
| Model | Quantization | Fits? | Approx tokens/sec |
|---|---|---|---|
| Llama 3.2 1B | Q4_K_M | Yes (0.8 GB) | ~144 tok/s |
| Llama 3.1 8B | Q4_K_M | Yes (~5 GB) | ~48 |
| Llama 3.1 8B | Q8_0 | Yes (~8 GB) | ~36 |
| Qwen 2.5 14B | Q4_K_M | Yes (~9 GB) | ~28 |
| Mistral Nemo 12B | Q4_K_M | Yes (~8 GB) | ~32 |
| Qwen 3 32B | Q4_K_M | Yes (~20 GB) | ~14 |
| DeepSeek R1 distilled 32B | Q4_K_M | Yes (~20 GB) | ~14 |
| Llama 3.1 70B | Q4_K_M | Yes (~42 GB) | ~8 |
| Llama 4 Scout 109B (MoE) | Q4_K_M | Yes (~64 GB) | ~6 |
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 M3 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
- NVIDIA RTX 3090 — 24GB VRAM, high-end consumer (used market)
- NVIDIA RTX 4090 — 24GB VRAM, flagship consumer
- NVIDIA RTX 5090 — 32GB VRAM, flagship consumer 2025+
- NVIDIA RTX 6000 Ada — 48GB VRAM, workstation
- NVIDIA A100 80GB — 80GB VRAM, data-center
- NVIDIA H100 80GB — 80GB VRAM, data-center flagship
- Apple M4 Max — 36GB VRAM, Apple Silicon flagship
- AMD Radeon RX 7900 XTX — 24GB VRAM, high-end AMD consumer
Sources
[1] apple.com/mac-studio · [2] mlx-lm benchmarks