selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

Best AI models for every VRAM size

Pick how much GPU memory you have. Each page ranks the strongest open-weight models that fit fully in it (chat, code, reasoning, vision, image generation, speech and embeddings), with the best quantisation and an estimated speed.

  1. 4GB1 101 models · 43 GPUsTop chat pickgranite-4.2-3bQ5_K_M · ~27 tok/s
  2. 6GB1 565 models · 21 GPUsTop chat pickgranite-4.2-3bQ8_0 · ~27 tok/s
  3. 8GB1 995 models · 86 GPUsTop chat pickQwen3-8BQ5_K_M · ~27 tok/s
  4. 10GB2 158 models · 7 GPUsTop chat pickQwen3-8BQ6_K · ~35 tok/s
  5. 12GB2 343 models · 32 GPUsTop chat pickQwen3-8BQ8_0 · ~27 tok/s
  6. 16GB2 439 models · 57 GPUsTop chat pickgpt-oss-20bQ8_0 · ~55 tok/s
  7. 20GB2 559 models · 4 GPUsTop chat pickLFM2.5-24B-A2BQ5_K_M · ~271 tok/s
  8. 24GB2 725 models · 27 GPUsTop chat pickOrnith-1.5-35B-A3BQ4_K_M · ~203 tok/s
  9. 32GB2 773 models · 30 GPUsTop chat pickOrnith-1.5-35B-A3BQ5_K_M · ~132 tok/s
  10. 48GB2 867 models · 16 GPUsTop chat pickOrnith-1.5-35B-A3BQ8_0 · ~133 tok/s
  11. 64GB2 901 models · 14 GPUsTop chat pickOrnith-1.5-35B-A3BQ8_0 · ~69 tok/s
  12. 80GB2 930 models · 5 GPUsTop chat pickGLM-4.5-AirQ4_K_M · ~203 tok/s
  13. 96GB2 941 models · 7 GPUsTop chat pickLaguna-S-2.1Q5_K_M · ~103 tok/s
  14. 128GB2 951 models · 12 GPUsTop chat pickLaguna-S-2.1Q6_K · ~59 tok/s

Memory figures include an 8K-token KV cache and runtime overhead. Speeds assume a typical card of each size; your GPU’s own page has estimates for its exact bandwidth.