Independent reference · October 2026
GPU VRAM Requirements Reference
Straight tables answering one question: how much video memory does a workload need? Figures come from official model cards (Ollama library, OpenAI Whisper, model GitHub repos) and are dated on every page.
LLMs
DeepSeek-R1, Llama 3.1, Qwen 2.5 — file size per variant and quantization (Q4 / Q8 / FP16).
Image generation
SD 1.5, SDXL, Flux and SD 3.5 tiers, plus LoRA training footprints.
Video generation
Wan 2.1 / 2.2, HunyuanVideo, LTX-Video — official VRAM figures.
Speech
Whisper model sizes and relative speeds, tiny through large-v3.
Card classes
Consumer vs datacenter GPU differences that matter beyond VRAM.
Quick conclusions
| Local hardware | What fits (rule of thumb) |
|---|---|
| 12GB card | 8B–14B models at Q4 quantization |
| 24GB card (RTX 3090 / 4090 class) | Up to ~32B models at Q4 (20GB file + overhead) |
| 48GB+ or 2×24GB | 70B models at Q4 (43GB file) |
Derived from the file-size tables on the LLM VRAM page. File size is the main part of the floor; runtime overhead adds a little on top.
Method notes
- LLM figures are model file sizes from the Ollama library, which is the dominant part of the VRAM floor — actual runtime usage runs slightly higher (context, activations).
- Image/video figures are official README requirements where available; community-sourced numbers are labeled as such.
- Nothing on this site is a benchmark; it is a requirements reference for picking hardware before you run anything.
- If your target model does not fit locally, renting a larger card by the hour is the usual alternative — rental marketplaces such as SpareHive list current-generation cards with full machine specs, which is what you match against these tables.