Mac miniAvailable
Mac mini M5 Pro
15-core CPU / 16-core GPU, 24GB
$1,699at Apple Store
24 GB · 307 GB/s
Check priceMemory guide · 16GB VRAM
What can you run on 16GB VRAM?
3 of the 16 open models we track fit in 16GB VRAM without spilling into system RAM, the biggest being GPT-OSS 20B. Speed depends on your card's memory bandwidth, so we show a range for 300–900 GB/s cards. Prices & stock checked 2026-10-23. Prices change daily and some retailers block our checker, so always confirm on the store page.
Models that fit in 16GB VRAM
| Model | Quant | Memory needed | Est. speed, 300–900 GB/s card |
|---|---|---|---|
| Qwen3.5 9Bdense · 9B active · chat, small agents | Q8_0 | ~12 GB | ~25–72 tok/s |
| Gemma 4 12Bdense · 12B active · chat, vision | Q4_K_M | ~9 GB | ~33–93 tok/s |
| GPT-OSS 20BMoE · 3.6B active · reasoning, tools | MXFP4 (native) | ~13 GB | ~98–198 tok/s |
We don't track any 16GB cards, so speed is shown for a typical 300–900 GB/s NVIDIA card (slower cards at the low end, faster ones at the high end). Fit assumes ~93% of VRAM is usable. How we estimate
Just out of reach: Gemma 4 26B-A4B (~18 GB at 4-bit), Qwen3.8 27B (~19 GB at 4-bit), GLM-4.7-Flash 30B-A3B (~21 GB at 4-bit).
Want more than 16GB?
We don't track any 16GB cards yet. These are the cheapest tracked rigs with more memory.
Strix HaloAvailable
GMKtec EVO-X2
Ryzen AI Max+ 395, 64GB / 1TB
$2,200at GMKtec store
64 GB · 256 GB/s
Check price
Mac StudioAvailable
Mac Studio M5 Max
18-core CPU / 32-core GPU, 36GB
$2,499at Apple Store
36 GB · 460 GB/s
Check priceFAQ
What's the biggest model I can run on 16GB VRAM?
Of the models we track, GPT-OSS 20B is the largest that fits fully in 16GB VRAM (MXFP4 (native), about 13 GB including runtime and ~8K context).
Can I run Llama 3.3 70B on 16GB VRAM?
Not fully in memory. At 4-bit it needs about 46 GB including runtime and ~8K context.
How are fit and speed estimated?
Memory needed = weights + 1.5 GB + 5% of weights. Decode speed ≈ 1 ÷ (active weights in GB ÷ (efficiency × memory bandwidth) + a small fixed per-token overhead), with efficiency and overhead fitted per platform to 43 public llama.cpp/MLX benchmarks (median error about 5%). They're estimates, not benchmarks.