How much VRAM does your AI model need?
VRAM decides which GPUs can run your model at all. Here is what popular open models for images, video, 3D, speech, and text need, and how to find the cheapest GPU that fits.
Before you compare GPUs on speed, check one number: VRAM, the memory on the GPU itself.
A model that fits in VRAM runs. A model that does not fit stops with an out-of-memory error, however fast the GPU is. So VRAM is the first filter, and it often shows that a smaller, cheaper GPU can do the work.
Here is what fills that memory, what popular open models need, and how to pick the GPU.
What fills the memory
The size of a model's weights is only the start. Five things add up:
- The weights. The parameters of the model, loaded onto the GPU.
- The other parts of the pipeline. Image and video models also carry a text encoder and a VAE. Black Forest Labs notes that the three parts of FLUX.2 [dev] do not all fit on an 80 GB H100 at once, so they take turns.
- The size of the output. Bigger images and longer clips take more. Stable Diffusion 3.5 Large goes from about 31.6 GiB at 1024 by 1024 to about 45.5 GiB at 2048 by 2048.
- The precision. Weights in 4-bit take far less room than weights in 16-bit, which is how FLUX.2 [dev] comes down to about 18 GB.
- How much runs at once. Larger batches and parallel requests need more room.
What popular open models need
These figures come from the maker of each model. Each one is for the setup the maker describes, so use it as the starting point for your own settings.
| Model | What it makes | VRAM |
|---|---|---|
| Whisper turbo | Text from speech | About 6 GB |
| FLUX.2 [klein] 4B | Images | About 8 GB |
| gpt-oss-20b | Text | 16 GB |
| FLUX.2 [dev], 4-bit with a remote text encoder | Images | About 18 GB |
| Wan 2.2 TI2V-5B | Video at 720p | 24 GB |
| Hunyuan3D-2.1, shape and texture | 3D assets | 29 GB |
| gpt-oss-120b | Text | 80 GB |
| Wan 2.2 T2V-A14B | Video at 480p, 720p | 80 GB |
| FLUX.2 [dev], every part in memory | Images | More than 80 GB |
One family can span a wide range. FLUX.2 runs from about 8 GB for [klein] 4B to more than 80 GB for [dev]. Wan 2.2 makes 720p video on an RTX 4090 with TI2V-5B, while its larger A14B models ask for 80 GB.
Which GPU fits
Line those figures up against the GPUs on the market, and most of the choice is already made:
| VRAM | GPU | What fits |
|---|---|---|
| 24 GB | RTX 4090 | Whisper, FLUX.2 [klein] 4B, gpt-oss-20b, FLUX.2 [dev] in 4-bit, Wan 2.2 TI2V-5B |
| 32 GB | RTX 5090 | All of the above, and Hunyuan3D-2.1 |
| 48 GB | L40S | All of the above, and FLUX.2 [dev] in 8-bit |
| 80 GB | A100 80GB, H100 SXM | All of the above, gpt-oss-120b, Wan 2.2 T2V-A14B, and FLUX.2 [dev] with its parts taking turns |
| 141 GB | H200 | All of the above, and FLUX.2 [dev] with every part in memory |
| 180 GB | B200 | All of the above, with room for more at once |
Same memory, different speed
Once your model fits on several GPUs, speed and rate decide. An A100 80GB and an H100 SXM hold the same 80 GB. On October 7, Uvik's index of list prices had the median A100 at $2.75 per GPU-hour and the median H100 at $4.20.
The H100 is faster and the A100 is cheaper per hour, so the better pick depends on your work. When you pay by the second, a faster GPU that finishes sooner can come out ahead. For a long batch that can run overnight, the lower rate is hard to beat.
Check it yourself
Start with the maker's figure on the model's page, and add room for your own resolution, frames, and batch size. Then rent the cheapest GPU with enough VRAM, and watch the real peak while your work runs:
nvidia-smi --query-gpu=memory.used,memory.total --format=csv -l 1
If the peak stays well under the total, you can go a size down next time. If it touches the total, go a size up or lower the precision. Moving parts of the pipeline to the CPU also frees VRAM, and it slows the work a little.
Where Texel fits
This check is what texel/auto is for. Texel is building the onchain GPU cloud on Robinhood Chain, with GPUs from an RTX 4090 to a B200, and texel/auto will read the VRAM your rental needs and pick the cheapest GPU that fits. You and your agent can skip the tables above.
If all you need is the output, Models, which opens first, skips the GPU altogether: send a prompt, get the output, and pay per output.
Sources
- Black Forest Labs, FLUX.2 and FLUX.2 dev with Diffusers.
- Wan Video, Wan 2.2.
- Tencent Hunyuan, Hunyuan3D-2.1.
- OpenAI, gpt-oss-120b and gpt-oss-20b model card and Whisper.
- AceCloud, VRAM requirements for image and video models, September 8, 2026.
- Uvik, GPU Price Index 2026, measured October 7, 2026.
