AI models that fit on a 192 GB GPU
Open models that fit in 192 GB of VRAM: 37 as published, 2 at FP8 and 17 at INT4. Each model is listed on the smallest tier it fits at that precision, so anything smaller is on the 141 GB page or below.
GPUs with 192 GB
Cards whose datasheet VRAM puts them in this tier, up to the next tier at 288 GB. Prices are the lowest live data-center on-demand rate per GPU.
| GPU | VRAM | From per GPU hour |
|---|---|---|
| AMD MI300X | 192 GB | $2.63/hr |
Fits as published
Models whose published weights, plus the flat overhead, fit this much memory with no quantization.
Text models
31 models.
Vision-language models
3 models.
| Model | Parameters | Published as | VRAM needed |
|---|---|---|---|
| Qwen2.5-VL-72B-Instruct | 73.4B | BF16 | 164 GB |
| InternVL3-78B | 78.4B | BF16 | 175 GB |
| Qwen3.8-Flash-Next-Uncensored-MLX | 71.3B | BF16 | 159 GB |
Image generation models
1 model.
| Model | Parameters | Published as | VRAM needed |
|---|---|---|---|
| Cosmos3-Super-Text2Image-4Step | 64.0B | BF16 | 143 GB |
Video generation models
1 model.
| Model | Parameters | Published as | VRAM needed |
|---|---|---|---|
| Cosmos3-Super-Image2Video | 64.6B | BF16 | 144 GB |
Fits at FP8
Models that fit only after quantizing the weights to 8 bits (1 byte per parameter). Needs an FP8 checkpoint or an engine that quantizes on load, and a GPU with FP8 support.
Text models
2 models.
| Model | Parameters | Published as | VRAM needed |
|---|---|---|---|
| Ling-3.0-flash | 127.5B | BF16 | 142 GB |
| xLAM-8x22b-r | 140.6B | BF16 | 157 GB |
Fits at INT4
Models that fit only after quantizing to 4 bits (0.5 byte per parameter). Needs a quantized checkpoint actually published for the model; check its Hugging Face page before relying on a row.
Text models
9 models.
| Model | Parameters | Published as | VRAM needed |
|---|---|---|---|
| MiMo-V2.5 | 310.8B | F8_E4M3 | 174 GB |
| MiMo-V2-Flash | 309.8B | F8_E4M3 | 173 GB |
| Hy3-preview | 298.8B | BF16 | 167 GB |
| Hy3-FP8 | 298.8B | F8_E4M3 | 167 GB |
| Hy3 | 298.8B | BF16 | 167 GB |
| GLM-5.3-Flash-FP8 | 321.3B | F8_E4M3 | 180 GB |
| GLM-5.3-Flash-Uncensored-FP8 | 321.3B | F8_E4M3 | 180 GB |
| IQuest-Q1 | 320.3B | BF16 | 179 GB |
| GLM-5.3-Flash | 321.3B | BF16 | 180 GB |
Vision-language models
6 models.
| Model | Parameters | Published as | VRAM needed |
|---|---|---|---|
| GLM-5.3-Flash | 321.3B | F8_E4M3 | 180 GB |
| Inkling-Small | 266.0B | BF16 | 149 GB |
| step3 | 321.0B | BF16 | 179 GB |
| GLM-5.3-Flash-BF16 | 321.3B | BF16 | 180 GB |
| apex-flash-1 | 321.3B | BF16 | 180 GB |
| apex-flash-1-abliterated | 321.3B | BF16 | 180 GB |
Other models
2 models.
| Model | Parameters | Published as | VRAM needed |
|---|---|---|---|
| GLM-5.3-Flash-UNCENSORED-FP8 | 321.3B | F8_E4M3 | 180 GB |
| GLM-5.3-Flash-ABLITERATED-FP8 | 321.3B | F8_E4M3 | 180 GB |
How these numbers are computed
Required VRAM is the weight size at each precision times a flat 1.2 overhead, the same figure every model page shows. It does not include a long context: the KV cache grows with every token, so a model near the top of a tier can need the next one at long context. Read how much VRAM you need for LLMs for the method, the VRAM and quantization glossary entries for the terms, and the VRAM calculator to size a model that is not listed.
Looking for a model by job rather than by memory? Start with coding, reasoning, chat and assistants, vision-language or see the full models directory.
