Local LLM on Apple M4 Ultra

Is the Apple M4 Ultra the right card for this?

Unified memory changes the arithmetic. There is no separate VRAM pool to fill, so the whole 128 GB is addressable by the model, and at 80 W this machine draws less than a single desktop GPU. The trade is bandwidth: 1000 GB/s is a fraction of what a data-centre part delivers, so very large models load and run rather than run fast. At about $4500, or roughly $35 per gigabyte of usable memory, it is the cheapest way to hold a model this size in one box. Catalogue throughput class: FP16 70B: 25-35, FP16 405B: 8-12.

At 2025 this is among the newest entries here, which cuts both ways. The memory technology is current, and the runtime support is the least settled of anything in this guide: check that the backend you intend to use has shipped a release naming this generation before buying for it.

What the price gap buys

No other card in this guide carries 128 GB, so there is no like-for-like price comparison to make. The nearest capacities are listed at the foot of this page, and moving to any of them changes which models fit, not just how fast they run.

At 1 year old this is still a current retail part, so the price above is a shop price rather than a listing price and stock is the usual constraint rather than condition. Buying new also means the 80 W figure is the number your power supply has to meet from day one.

Where the Apple M4 Ultra sits in this guide

Across the 20 cards catalogued here, the Apple M4 Ultra ranks 6th on memory bandwidth at 1000 GB/s, 16th cheapest at about $4500, and 2nd on bandwidth per watt at 12.5 GB/s per watt. Those three positions, not the capacity figure, are what separate it from other 128 GB parts.

Unified LPDDR5X is laptop-class memory shared between CPU and GPU, so there is no copy step between "system RAM" and "video RAM" and no fixed split between them. 1000 GB/s is modest next to discrete graphics memory, but the model never has to be moved, and the whole 128 GB is available to it.

On Apple hardware the fast path is MLX, Apple's own array framework, with Ollama as the simpler alternative running through Metal. Our card records for Apple parts cite mlx-lm benchmarks for exactly this reason: the same machine gives materially different numbers depending on which of the two you use.

Which catalogue models fit on Apple M4 Ultra?

Of the 46 models in our catalogue with per-size memory figures, 41 have at least one size that fits in 128 GB at Q4_K_M with room for context. The largest of each are below, biggest first.

Model	Largest size that fits	Weights at Q4_K_M
Mistral Large	123B	70 GB
GPT-OSS	120B	70 GB
Llama 3.2 Vision	90B	55 GB
Qwen 2.5	72B	44 GB
Llama 3.1	70B	42 GB
Llama 3.3	70B	42 GB
DeepSeek R1	70B	42 GB
Code Llama	70B	42 GB

A further 33 smaller models also fit; see the model catalogue.

The next sizes up are out of reach without offloading: Qwen 3 235B at 130 GB, minimax M2 230B at 130 GB, minimax M2.1 230B at 130 GB.

Expected tokens per second on Apple M4 Ultra

Memory bandwidth, not compute, is the bottleneck for single-stream token generation. The Apple M4 Ultra's 1000 GB/s puts the following ceiling on a batch of one at 2048 tokens of context.

Model	Quantization	Status	Approx tokens/sec
Llama 3.2 1B	Q4_K_M	Fits in VRAM	~180 tok/s
Llama 3.1 8B	Q4_K_M	Fits in VRAM	~60 tok/s
Llama 3.1 8B	Q8_0	Fits in VRAM	~45 tok/s
Mistral Nemo 12B	Q4_K_M	Fits in VRAM	~40 tok/s
Qwen 2.5 14B	Q4_K_M	Fits in VRAM	~35 tok/s
Qwen 3 32B	Q4_K_M	Fits in VRAM	~18 tok/s
DeepSeek R1 distilled 32B	Q4_K_M	Fits in VRAM	~18 tok/s
Llama 3.1 70B	Q4_K_M	Fits in VRAM	~10 tok/s
Llama 4 Scout 109B (MoE)	Q4_K_M	Fits in VRAM	~8 tok/s

These come from a bandwidth heuristic (tokens/sec ≈ bandwidth in GB/s × a per-model efficiency factor), not from a test rig in our office. Real numbers move with model architecture, batch size and KV-cache size. Treat them as an order of magnitude, and treat the "reported throughput class" row in the table above as the community-measured figure.

Build recommendations for Apple M4 Ultra

Best model to download first

Llama 3.1 8B to check the stack works, then the largest model in the table above that you actually need. At 128 GB the constraint stops being capacity. 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.

Sources

[1] apple.com/mac-studio · VRAM figures for catalogue models from src/data/models.json (2026-06-29); card specifications from src/data/gpus.json (2026-06-29).