NVIDIA · NVIDIA GPUs · 2025
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)
Researched from manufacturer data and public sources; not tested by us.· Last reviewed Oct 11, 2026
Popular models at Q4_K_M with a 32K-token context. Memory needed from real model files; speed is the theoretical ceiling.
Model weightsContext cache (KV)RuntimeUsable limit (colour = verdict)
| Model | Verdict | Memory | Speed ceiling |
|---|---|---|---|
| Qwen3 1.7B | Runs | ≤ 1,614 tok/s | |
| Qwen3.5 9B | Runs | ≤ 315 tok/s | |
| Gemma 4 12B | Runs | ≤ 252 tok/s | |
| gpt-oss-20b | Runs | ≤ 895 tok/s | |
| Qwen3.6 27B | Runs | ≤ 107 tok/s | |
| Qwen3.6 35B A3B | Runs | ≤ 972 tok/s | |
| Llama 3.3 70B | Runs | ≤ 42 tok/s | |
| gpt-oss-120b | Runs | ≤ 654 tok/s | |
| Qwen3.5 122B A10B | Runs | ≤ 293 tok/s | |
| Qwen3.5 397B A17B | Partly in system RAM | — |
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.
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.
A 300 W Max-Q edition with the same 96 GB also exists (somewhat slower, easier to cool).
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
Best for: Serious users who want 100B-class models at full GPU speed in a normal tower.
If budget is no object, it is the fastest single-card way to run 100B-class models at home.
Source: Manufacturer specifications
How well the popular local-AI tools run on it today (editorial assessment).
Measured results with their source. We never invent speeds.
No measured results yet. The table above shows the theoretical ceiling instead.
Apple silicon
Largest popular model at 4-bit: Qwen3.5 122B A10B
128 GB at 614 GB/s: gpt-oss-120b and 122B MoE models at a good pace, almost silently.
Ryzen AI
Largest popular model at 4-bit: Qwen3.5 122B A10B
128 GB of unified memory in a mini PC for less than a Mac Studio: the homelab favourite for MoE models.
DGX Spark / GB10
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.