NVIDIA GPUs
NVIDIA GeForce RTX 5090
- Memory
- 32 GB
- Bandwidth
- 1.79 TB/s
Largest popular model at 4-bit: Qwen3.6 35B A3B
The fastest consumer card for local AI: 32 GB and 1.8 TB/s make 30B-class models fly.
from $4,329 (typical price)Recommended
Apple · Apple silicon · 2026
Apple Mac Studio (M5 Ultra)
Illustration
Recommended
Up to 512 GB at 1.2 TB/s: the only single box that holds DeepSeek R1 or Qwen3.5 397B at 4-bit.
- Memory
- 96 GB–512 GB
- Bandwidth
- 1.2 TB/s
- Power
- 480 W
- Largest popular model at 4-bit
- Qwen3.5 397B A17B
from $5,499 (launch price)
Researched from manufacturer data and public sources; not tested by us.· Last reviewed Oct 11, 2026
What it can run
EstimatePopular models at Q4_K_M with a 32K-token context. Memory needed from real model files; speed is the theoretical ceiling.
Change model size or context
Model weightsContext cache (KV)RuntimeUsable limit (colour = verdict)
| Model | Verdict | Memory | Speed ceiling |
|---|---|---|---|
| Qwen3 1.7B | Runs | ≤ 1,081 tok/s | |
| Qwen3.5 9B | Runs | ≤ 211 tok/s | |
| Gemma 4 12B | Runs | ≤ 169 tok/s | |
| gpt-oss-20b | Runs | ≤ 600 tok/s | |
| Qwen3.6 27B | Runs | ≤ 71 tok/s | |
| Qwen3.6 35B A3B | Runs | ≤ 651 tok/s | |
| Llama 3.3 70B | Runs | ≤ 28 tok/s | |
| gpt-oss-120b | Runs, tight | ≤ 438 tok/s | |
| Qwen3.5 122B A10B | Runs | ≤ 196 tok/s | |
| Qwen3.5 397B A17B | Runs | ≤ 117 tok/s |
Theoretical maximum for generating text: memory bandwidth divided by the bytes read per token. Real runs usually reach about half to three quarters of it.
Where to buy
Prices are editorial references (launch price and a typical street price on the date shown), not live prices. In 2026 the memory shortage has pushed many prices well above launch.
- Mac Studio M5 Ultra, 96 GB96 GBLaunch price: $5,499
- Check price Amazon.com (opens in a new tab)Amazon.comCheck availability
Other regions (3)
- Check price Amazon.co.uk (opens in a new tab)Amazon.co.ukUnited Kingdom · Check availability · Search results
- Check price Amazon.es (opens in a new tab)Amazon.esSpain · Check availability · Search results
- Check price Amazon.de (opens in a new tab)Amazon.deGermany · Check availability · Search results
- Mac Studio M5 Ultra, 256 GB256 GB
- Check price Amazon.com (opens in a new tab)Amazon.comCheck availability · Search results
Other regions (3)
- Check price Amazon.co.uk (opens in a new tab)Amazon.co.ukUnited Kingdom · Check availability · Search results
- Check price Amazon.es (opens in a new tab)Amazon.esSpain · Check availability · Search results
- Check price Amazon.de (opens in a new tab)Amazon.deGermany · Check availability · Search results
- Mac Studio M5 Ultra, 512 GB512 GB
Apple announced the 512 GB configuration for late October 2026.
- Check price Amazon.com (opens in a new tab)Amazon.comCheck availability · Search results
Other regions (3)
- Check price Amazon.co.uk (opens in a new tab)Amazon.co.ukUnited Kingdom · Check availability · Search results
- Check price Amazon.es (opens in a new tab)Amazon.esSpain · Check availability · Search results
- Check price Amazon.de (opens in a new tab)Amazon.deGermany · Check availability · Search results
Prices are shown only if updated in the last 24 hours and may change; the retailer’s price at checkout applies. We may earn a commission from these links at no cost to you. Affiliate disclosure
Overview
Editorial opinionBest for: Running the largest open models at home without a rack of GPUs.
Editorial opinionWould we buy it today?
If you need 200B+ models locally, nothing else is this simple. Otherwise the M5 Max is better value.
Pros and cons
Editorial opinionPros
- Up to 512 GB of unified memory
- 1.2 TB/s bandwidth
- Quiet, under 500 W
Cons
- Very expensive at 256–512 GB
- Prompt processing on huge models takes a while
Good to know
macOS lets the GPU use about two thirds of memory on machines with up to 36 GB, and about three quarters above that. The limit can be raised (sudo sysctl iogpu.wired_limit_mb=…), leaving less for macOS and other apps; our figures use the default.
Specifications
Factual data- Manufacturer
- Apple
- Chip
- M5 Ultra
- Architecture
- Apple silicon
- Type
- Unified-memory computer
- Form factor
- Desktop
- Released
- 2026
- Availability
- On sale
- Memory options
- 96 GB / 256 GB / 512 GB
- Memory type
- Unified LPDDR
- Memory bandwidth
- 1.2 TB/s
- Usable for models (default)
- 384 GB / 512 GB
- Board / max power
- 480 W
- Compute stack
- Metal / MLX
- Models
- Any model that fits
- Several units for one model
- Yes
Source: Manufacturer specifications
Software support
Editorial opinionHow well the popular local-AI tools run on it today (editorial assessment).
- llama.cppExcellent
- OllamaExcellent
- LM StudioExcellent
- vLLMPoor
- MLXExcellent
Measured speeds
Factual dataMeasured results with their source. We never invent speeds.
No measured results yet. The table above shows the theoretical ceiling instead.
Alternatives
DGX Spark / GB10
NVIDIA GB10 Grace Blackwell (DGX Spark and partners)
- Memory
- 128 GB
- Bandwidth
- 273 GB/s
Largest popular model at 4-bit: Qwen3.5 122B A10B
128 GB with CUDA in a tiny box: the NVIDIA answer to Strix Halo and the Mac.
from $5,000 (typical price)Recommended
NVIDIA GPUs
NVIDIA RTX PRO 6000 Blackwell
- Memory
- 96 GB
- Bandwidth
- 1.79 TB/s
Largest popular model at 4-bit: Qwen3.5 122B A10B
96 GB on one card at RTX 5090 speed: gpt-oss-120b and 122B MoE models entirely in GPU memory.
from $8,565 (typical price)For advanced users