Which open models fit your GPU
How much graphics memory each open-weight model in our release tracker needs to run locally, at 16-bit, 8-bit and 4-bit quantization, and which fit on 8, 12, 16, 24, 48, 80 and 128 GB cards. These are estimates from each model's parameter count, not measured runs.
By memory size
Largest model that fits, and how many fit| VRAM | 16-bit | 8-bit | 4-bit |
|---|---|---|---|
| 8 GB | CohereLabs tiny-aya-l2-thinker (3.35B), 135 models fit | nvidia NV-Reason-CT (5.32B), 177 models fit | google translategemma-12b-it (13.2B), 236 models fit |
| 12 GB | google translategemma-4b-it (4.97B), 173 models fit | Qwen Qwen3.5-9B-Base (9.65B), 225 models fit | microsoft Phi-4-reasoning-vision-15B (15.1B), 248 models fit |
| 16 GB | nvidia NV-Reason-CT (5.32B), 177 models fit | google translategemma-12b-it (13.2B), 236 models fit | google gemma-4-26B-A4B-it-qat-q4_0-unquantized (26.5B), 256 models fit |
| 24 GB | Qwen Qwen3.5-9B-Base (9.65B), 225 models fit | microsoft Phi-4-reasoning-vision-15B (15.1B), 248 models fit | Qwen Qwen3.6-35B-A3B (36B), 287 models fit |
| 48 GB | microsoft Phi-4-reasoning-vision-15B (15.1B), 248 models fit | Qwen Qwen3.6-35B-A3B (36B), 287 models fit | Qwen Qwen3-Coder-Next-Base (79.7B), 293 models fit |
| 80 GB | nvidia Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 (33B), 281 models fit | nvidia Nemotron-Labs-TwoTower-30B-A3B-Base-BF16 (63.2B), 289 models fit | Qwen Qwen3.5-122B-A10B (125B), 297 models fit |
| 128 GB | ai21labs AI21-Jamba2-Mini (51.6B), 288 models fit | nvidia gpt-oss-puzzle-88B (90.8B), 294 models fit | Qwen Qwen3.8-Flash-Next (180B), 298 models fit |
Every model
Estimated GB, smallest model firstModels whose weight files declare no parameter count are left out. Each name links to the official model card.
MethodHow these figures are worked out
Memory needed = parameter count x bytes per weight x 1.2. Bytes per weight: 2 at 16-bit, 1 at 8-bit, 0.5 at 4-bit. The extra 20% is our allowance for the KV cache at a short context and runtime buffers; long contexts, large batches and runtimes that keep some layers at higher precision need more, and common 4-bit files (such as GGUF Q4_K_M) use slightly more than 4 bits per weight. Mixture-of-experts models are counted at their total parameters, since all experts must sit in memory. GB here is 109 bytes. A model fits a tier when the estimate is at or below the card's memory; leave headroom for the display and other programs.
Parameter counts come from our open model release tracker, read from each publisher's Hugging Face repository. More on how we source figures: how we collect our data.