What can you run on a 32GB graphics card?
With 32GB VRAM you can run 22 of the 51 open-source models we track — up to about 36B parameters at Q4 quantization. The most popular that fits is Qwen3.8-27B, needing 19 GB and generating around ~71 tok/s on typical hardware of this size.
Typical machines: RTX 5090. After the operating system takes its share, about 33 GB is free for a model.
Having the memory is not the same as having the speed
32GB VRAM decides whether a model loads. How fast it answers is decided by memory bandwidth — how quickly your machine can read the model, which it must do once for every word it writes. The table quotes a RTX 5090; pick yours below to see the difference. Why bandwidth and not the processor →
Models that run on 32GB VRAM in 2026
| Model↕ | Size↕ | Memory↑ | Est. speed↕ | Downloads↕ | Feels like |
|---|---|---|---|---|---|
| LFM2.5-230MLiquid Ai | 230M | 808 MBRuns well | ~7593 tok/s | 58,011 | Fast |
| MiniCPM5-1BOpenBMB | 1.1B | 1.3 GBRuns well | ~1785 tok/s | 667,154 | Fast |
| MiniCPM5-2BOpenBMB | 2.5B | 2.3 GBRuns well | ~746 tok/s | 206,774 | Fast |
| LFM2.5-2.6BLiquid Ai | 2.7B | 2.5 GBRuns well | ~696 tok/s | 98,713 | Fast |
| LFM2.5-VL-3BLiquid Ai | 3.1B | 2.5 GBRuns well | ~696 tok/s | 31,296 | Fast |
| North-Micro-Vision-InstructCohere | 2.5B | 2.5 GBRuns well | ~776 tok/s | 120,354 | Fast |
| granite-4.2-3bIbm | 3.7B | 3.1 GBRuns well | ~519 tok/s | 32,017 | Fast |
| Mage-VLMicrosoft | 4.7B | 3.9 GBRuns well | ~429 tok/s | 29,690 | Fast |
| LFM2.5-8B-A1BLiquid Ai · MoE | 8.5B | 5.9 GBRuns well | ~1913 tok/s1B active | 80,632 | Fast |
| Ling-3.0-tinyInclusionai · MoE | 7.9B | 6.2 GBRuns well | ~1207 tok/s1.6B active | 25,727 | Fast |
| granite-4.2-8bIbm | 8.8B | 6.6 GBRuns well | ~218 tok/s | 52,946 | Fast |
| Ornith-1.5-9BDeepreinforce | 9.7B | 6.9 GBRuns well | ~202 tok/s | 453,339 | Fast |
| gemma-4-12B-itGoogle DeepMind | 12B | 9.6 GBRuns well | ~161 tok/s | 2,958,440 | Fast |
| Qwen3.8-27BAlibaba / Qwen | 28B | 19 GBRuns well | ~71 tok/s | 7,703,400 | Fast |
| diffusiongemma-26B-A4B-itGoogle DeepMind · MoE | 26B | 19 GBRuns well | ~447 tok/s4B active | 708,393 | Fast |
| Muse-Glimmer-30BMeta AI | 30B | 20 GBRuns well | ~65 tok/s | 441,881 | Fast |
| granite-4.2-30bIbm | 29B | 21 GBRuns well | ~66 tok/s | 17,894 | Fast |
| North-Mini-Code-1.0Cohere · MoE | 30B | 22 GBRuns well | ~303 tok/s6.1B active | 11,742 | Fast |
| Laguna-XS-2.1Poolside · MoE | 33B | 23 GBRuns well | ~288 tok/s6.7B active | 32,445 | Fast |
| Nex-N2.5-miniNex Agi · MoE | 35B | 24 GBRuns well | ~275 tok/s7B active | 4,543 | Fast |
| Ornith-1.5-35B-A3BDeepreinforce · MoE | 36B | 24 GBRuns well | ~643 tok/s3B active | 359,808 | Fast |
| NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4NVIDIA · MoE | 18B | 28 GBTight | ~275 tok/s3B active | 1,261,605 | Fast |
Sorted by memory, smallest first — click any column to change it.
All figures at Q4_K_M quantization, 4K context. “Memory” includes the model, its context cache and runtime overhead. Speed is calculated, not benchmarked: 1792 GB/s × 0.65 ÷ bytes read per token, assuming a RTX 5090. Scale it by your own machine's bandwidth from the list above. Mixture-of-experts models read only their active parameters per token, which is why some large ones outrun smaller dense models.
What you'd need more memory for
The next models up, and what they ask for.