The GPU a model needs comes down to two numbers: how large its checkpoint is on disk, and how much room its KV cache needs for the context length and the number of requests running at once. This hub lines up the models worth self-hosting today against the smallest DeployGPU GPU class that fits each one, calculated rather than measured, so you can pick a card before reading a single benchmark.
How to size a GPU for a model
Weights come first, and the format a model ships in matters more than its parameter count alone. gpt-oss stores its mixture-of-experts weights at 4.25 bits per parameter in MXFP4, so 117 billion parameters fit in 65.2 GB rather than the 234 GB that plain 16-bit precision would need. Gemma 4 and Qwen ship in BF16 at 2 bytes per parameter, and official FP8 or NVFP4 builds exist for some of them. After the weights, the KV cache adds a cost that grows with context length and concurrency, and that cost behaves differently for a plain transformer than for a model that mixes full attention with sliding-window or linear-attention layers. The KV cache explainer works through the formula with worked examples.
Which GPU fits which model
Every row below is a native-precision checkpoint: BF16 for Gemma 4, Qwen, Llama and the 30B models, the native MXFP4 build for gpt-oss and Kimi K3, and the FP8 releases of DeepSeek V4 and GLM-5.3. Quantized and FP8 builds shrink the floor further, and each model's own page covers that range.
| Model | Weights (calculated) | Smallest DeployGPU class |
|---|---|---|
| gpt-oss-20b | 13.8 GB, MXFP4 | 48 GB |
| Gemma 4 E4B | 16.0 GB, BF16 | 48 GB |
| Qwen3.5-9B | 19.3 GB, BF16 | 48 GB |
| Gemma 4 12B | 23.9 GB, BF16 | 48 GB |
| Gemma 4 26B-A4B | 51.6 GB, BF16 | 80 GB |
| Qwen3.8-27B | 55.6 GB, BF16 | 80 GB |
| Gemma 4 31B | 62.5 GB, BF16 | 80 GB |
| gpt-oss-120b | 65.2 GB, MXFP4 | 80 GB |
| Qwen3.6-35B-A3B | 71.9 GB, BF16 | 96 GB |
| Llama 3.1 8B | 16.1 GB, BF16 | 48 GB |
| Muse Glimmer 30B | 59.6 GB, BF16 | 80 GB |
| Qwen3-Coder 30B-A3B | 61.1 GB, BF16 | 80 GB |
| GLM-4.7-Flash | 62.4 GB, BF16 | 80 GB |
| Nemotron 3.5 Lightning 30B-A3B | 65.8 GB, BF16 | 80 GB |
| Llama 3.3 70B | 141.1 GB, BF16 | Four 80 GB GPUs, or two H200s |
| DeepSeek V4-Flash | 166.9 GB, FP8 and FP4 | Two H200s, four for room to serve |
| Llama 4 Scout | 217.3 GB, BF16 | Four 80 GB GPUs, or two H200s |
| GLM-5.3-Flash | 328.3 GB, FP8 | Four H200s, or eight H100s |
| GLM-5.3 | 755.6 GB, FP8 | Eight H200s |
| DeepSeek V4-Pro | 892.7 GB, FP8 and FP4 | Eight H200s |
| Kimi K3 | 1,560.9 GB, MXFP4 | Eight B300s |
Figures are calculated from each model's published safetensors files, not measured on a live rental. A checkpoint that just barely fits a GPU leaves little room for the KV cache, so a model landing at the top of a tier is often more comfortable one tier up.
The models that need a whole 8-GPU machine are compared on what it takes to self-host frontier models. Image and video models size differently, because a pipeline loads a text encoder and a VAE next to the main model: the FLUX and Wan 2.2 pages cover them. Training needs far more memory than serving the same model, and GPU memory for fine-tuning has that math, while the quantization guide explains which smaller formats run on which GPU.
The pages below cover the sizing math, the license and a command to start from for each family.