Pick an open-source AI model and this calculator tells you how much GPU memory (VRAM) you need, whether it runs on your own PC or Mac, the cheapest GPU servers that can host it, and what the same usage would cost through a pay-per-token API. It covers Gemma 4, Qwen, Llama, DeepSeek, Mistral and gpt-oss models.
1. Run it locally
Which home or office hardware can hold this model?
3. Or use an API
2. Rent a GPU server
Cheapest hosting options that fit, sorted by estimated monthly cost at your hours per day. Monthly plans are fixed-price dedicated servers or VPS.
| Provider | GPU | VRAM | Price | Est. per month |
|---|
Quick start
Estimates only. VRAM = model weights + KV cache for your context length + runtime overhead. Real usage varies with the inference engine, batch size and number of users. Hosting prices checked September 2026; API prices from OpenRouter (cheapest listed provider, per 1M tokens) on 24 September 2026. € prices are sorted at about $1.17 per €.
How the Calculator Works
- Model weights: parameters × bytes per parameter. 4-bit is about 0.56 GB per billion parameters, 8-bit about 1.06 GB and 16-bit about 2 GB. Mixture-of-Experts (MoE) models need memory for all experts, not just the active ones.
- KV cache: memory that grows with context length. Longer chats, documents and agents need more.
- Overhead: CUDA context, activations and fragmentation, about 1 GB plus 3%.
- Fit: we assume about 92% of a GPU’s VRAM is usable. “Tight” means it should load but leaves little room for more users or longer context. Macs can use roughly 70–75% of their unified memory for the GPU.
- Costs: hourly GPUs are multiplied by your hours per day × 30.4 days; monthly plans are fixed. API cost uses your daily token usage × 30.4.
Self-Host or Use an API?
For light or occasional use, an API is almost always cheaper: you only pay for the tokens you use. Self-hosting on a GPU server wins when you have heavy, steady traffic, need data privacy (prompts never leave your server), want to run a fine-tuned or uncensored model, or need a model that no API offers. Many teams start with an API and move to a dedicated GPU once usage grows.
FAQ
Roughly 0.6 GB per billion parameters in 4-bit, plus a few GB for context. An 8B model needs about 6 GB, a 32B model about 20 GB and a 70B model about 42 GB.
Yes, in 4-bit on a 48 GB GPU such as an RTX A6000, A40 or L40S, or in 8-bit on an 80 GB A100 or H100.
No. A Mixture-of-Experts model is faster because fewer parameters are active per token, but all experts must still fit in memory.
For most chat and coding use, modern 4-bit formats (Q4_K_M, AWQ, NVFP4, MXFP4) lose very little quality. Use 8-bit when accuracy matters most.
Yes. Apple Silicon Macs share memory between CPU and GPU, so a 64–128 GB Mac can run models that need a data-center GPU on a PC, though usually more slowly.
Want the details for a specific model? See our guides to Qwen, DeepSeek, Gemma 4, Mistral, Llama and Ollama hosting.