Before you rent a GPU for a model, check it fits. Each table shows how much GPU memory a popular open model needs and the cheapest cloud machine that runs it right now. The most downloaded models from Hugging Face come first.
Chat and text LLMs
Llama, Qwen, gpt-oss, Mistral, DeepSeek and more, served with vLLM.
| Model | GPU memory needed | From /h |
|---|---|---|
| Qwen3-0.6B Qwen | 4.9 GB | $0.071 |
| gpt2 openai-community | 2.6 GB | $0.071 |
| Qwen3-14B Qwen | 37 GB | $0.335 |
| Qwen3-4B Qwen | 13 GB | $0.130 |
| Qwen3-8B Qwen | 22 GB | $0.207 |
| gpt-oss-20b openai | 18 GB | $0.175 |
| Qwen2.5-7B-Instruct Qwen | 20 GB | $0.175 |
| DeepSeek-V3.2 deepseek-ai | 784 GB | $83.44 |
| Qwen2.5-0.5B-Instruct Qwen | 3.4 GB | $0.071 |
| Llama-3.1-8B-Instruct meta-llama · licence approval needed | 21 GB | $0.207 |
| dolphin-2.9.1-yi-1.5-34b dphn | 81 GB | $0.975 |
| Qwen2.5-1.5B-Instruct Qwen | 5.9 GB | $0.071 |
| Qwen3.6-35B-A3B-NVFP4 nvidia | 29 GB | $0.271 |
| Qwen3-1.7B Qwen | 7.8 GB | $0.071 |
| DeepSeek-V4-Flash-0731 deepseek-ai | 189 GB | $3.95 |
| Qwen2.5-3B-Instruct Qwen | 9.4 GB | $0.071 |
| pythia-160m EleutherAI | 2.7 GB | $0.071 |
| Llama-3.2-1B-Instruct meta-llama · licence approval needed | 5.3 GB | $0.071 |
| Qwen3-32B Qwen | 77 GB | $0.975 |
| gpt-oss-120b openai | 75 GB | $0.975 |
Vision-language models
Models that read images and text, such as Gemma and Qwen-VL.
| Model | GPU memory needed | From /h |
|---|---|---|
| gemma-4-31B-it | 81 GB | $0.975 |
| Qwen3.8-27B Qwen | 66 GB | $0.975 |
| gemma-4-26B-A4B-it | 62 GB | $0.975 |
| Qwen3.5-4B Qwen | 14 GB | $0.130 |
| GLM-5.3-Flash zai-org | 374 GB | $10.21 |
| Qwen3.5-9B Qwen | 25 GB | $0.271 |
| Qwen3-VL-8B-Instruct Qwen | 23 GB | $0.234 |
| Qwen2.5-VL-7B-Instruct Qwen | 21 GB | $0.207 |
Speech to text
Whisper and other transcription models.
| Model | GPU memory needed | From /h |
|---|---|---|
| whisper-large-v3-turbo openai | 3.9 GB | $0.071 |
| whisper-small openai | 2.6 GB | $0.071 |
| whisper-large-v3 openai | 13 GB | $0.130 |
| whisper-base openai | 2.2 GB | $0.071 |
| Qwen3-ASR-1.7B Qwen | 7.6 GB | $0.071 |
| Qwen3-ASR-0.6B Qwen | 4.3 GB | $0.071 |
Image generation
Stable Diffusion, SDXL and FLUX.
| Model | GPU memory needed | From /h |
|---|---|---|
| stable-diffusion-xl-base-1.0 stabilityai | 11 GB | $0.071 |
| stable-diffusion-v1-5 stable-diffusion-v1-5 | 6.7 GB | $0.071 |
| RealVisXL_V5.0 SG161222 | 11 GB | $0.071 |
| dreamshaper-7 Lykon | 6.7 GB | $0.071 |
| Z-Image-Turbo Tongyi-MAI | 25 GB | $0.516 |
| animagine-xl-4.0 cagliostrolab | 11 GB | $0.071 |
| sd-turbo stabilityai | 6.6 GB | $0.071 |
| sdxl-turbo stabilityai | 11 GB | $0.071 |
Embeddings
Models that turn text into vectors for search and RAG.
| Model | GPU memory needed | From /h |
|---|---|---|
| bge-small-en-v1.5 BAAI | 1.6 GB | $0.071 |
| bge-large-en-v1.5 BAAI | 2.3 GB | $0.071 |
| bge-base-en-v1.5 BAAI | 1.8 GB | $0.071 |
| Qwen3-Embedding-0.6B Qwen | 2.9 GB | $0.071 |
| multilingual-e5-large intfloat | 2.8 GB | $0.071 |
| granite-embedding-small-english-r2 ibm-granite | 1.6 GB | $0.071 |
GPU memory to serve each model at its published precision with room for the KV cache. Machines and prices from GridPup, updated through the day.
Next steps
Deploy a model as an API
Serve any model here behind an https API with keys, on the cheapest GPU that fits.
Benchmark it on real GPUs
Tokens per second and users per GPU, so you know which GPU to rent before you deploy.
Fine-tune it
Train or fine-tune an open model on a cloud GPU from your browser.
Frequently asked questions
How much VRAM do I need to run an LLM?
Roughly 2 GB per billion parameters at 16-bit precision for the weights, plus room for the KV cache that holds the conversation. A 7B-8B model needs about 20 GB to serve with vLLM, a 32B model about 75 GB, and 4-bit (AWQ) versions need about half as much. The tables on this page show the figure for each model.
What is the cheapest GPU to run an 8B model?
An 8B model at 16-bit precision needs about 20-22 GB, so one 24 GB GPU (RTX 3090, RTX 4090, L4, A5000) or two 12 GB GPUs run it. A 4-bit AWQ version fits in about 10 GB, on a single RTX 3060. The tables show the cheapest machine that fits each model right now.
Can I run a model across several GPUs?
Yes. vLLM splits a model across the GPUs of one machine, so a model that needs 75 GB runs on one 80 GB GPU or on four 24 GB GPUs. That is why the cheapest machine for a model is sometimes several small GPUs.
Where does this data come from?
From GridPup (gridpup.com), which reads open models from Hugging Face, works out the GPU memory each needs, and matches it with live GPU offers from cloud providers. Open a model on GridPup to see every GPU that fits and deploy it as an API.