What can you run on a 24GB graphics card?
With 24GB VRAM you can run 24 of the 66 open-source models we track, up to 35B parameters, using each model's Q4 file where one is published and otherwise the smallest file that is. The most downloaded that fits is Qwen3.8-27B, needing 19 GB.
Typical machines: RTX 4090 · RTX 3090 · RX 7900 XTX. We hold back up to 1.5GB of the card for the display, leaving 24 GB for a model.
24GB VRAM decides whether a model loads. How fast it answers depends on your machine's memory bandwidth, which differs between machines with the same amount of memory. Why bandwidth sets the speed →
Models that run on 24GB VRAM in 2026
| Model↕ | Size↕ | Memory↑ | Downloads (30 days)↕ |
|---|---|---|---|
| MiniCPM-V-4.7-1BOpenBMB | 1.6B | 1.4 GBRuns well | 184,495 |
| K2-Horizon-0.9BMBZUAI Institute of Foundation Models | 1.1B | 1.4 GBRuns well | 13,425 |
| LFM3-1.7BLiquid AI | 1.7B | 1.8 GBRuns well | 15,885 |
| LFM3-VL-2BLiquid AI | 1.9B | 1.8 GBRuns well | 4,193 |
| MiniCPM5-2BOpenBMB | 2.5B | 2.3 GBRuns well | 1,180,516 |
| d1-3BLiquid AI | 3.1B | 2.5 GBRuns well | 31,374 |
| granite-4.2-3bIBM | 3.7B | 3.1 GBRuns well | 37,116 |
| Mage-VLMicrosoft | 4.7B | 3.9 GBRuns well | 7,645 |
| FrogNano-4B-2609Microsoft | 4.7B | 3.9 GBRuns well | 2,663 |
| K2-Horizon-3.7BMBZUAI Institute of Foundation Models | 5.1B | 4.3 GBRuns well | 24,267 |
| North-Micro-Vision-InstructCohere | 2.5B | 6.0 GBRuns well | 33,984 |
| Ling-3.0-tinyinclusionAI · MoE | 7.9B1.6B active | 6.2 GBRuns well | 13,391 |
| granite-4.2-8bIBM | 8.8B | 6.6 GBRuns well | 70,118 |
| K2-Horizon-7BMBZUAI Institute of Foundation Models | 9B | 6.8 GBRuns well | 87,218 |
| Ornith-1.5-9BDeepReinforce | 9.7B | 6.9 GBRuns well | 462,696 |
| MiMo-V2.6-Distill-Qwen-9BXiaomi | 9.4B | 7.0 GBRuns well | 19,920 |
| granite-switch-4.1-8bIBM | 8.8B | 7.8 GBRuns well | 3,734 |
| Qwen3.8-27BAlibaba / Qwen | 28B | 19 GBRuns well | 7,509,924 |
| Hy-Memory-30B-A3BTencent · MoE | 30B3B active | 20 GBRuns well | 5,468 |
| granite-4.2-30bIBM | 29B | 21 GBTight | 36,633 |
| Muse-Glimmer-30BMeta AI | 30B | 21 GBTight | 202,402 |
| Xing4.0-29B-A4BChina Telecom AI (TeleAI) · MoE | 31B4B active | 23 GBTight | 7,749 |
| MiniCPM-V-4.7-35B-A3BOpenBMB · MoE | 35B3B active | 24 GBTight | 334,870 |
| Nex-N2.5-miniNex-AGI · MoE | 35B7B active | 24 GBTight | 3,200 |
Sorted by memory, smallest first. Click any column to change it.
Step up to 32GB VRAM and 6 more models come within reach
Including NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 (32B), Ornith-1.5-35B-A3B (36B) and K2-Horizon-MoVA-36B-A4B (37B).
See everything 32GB VRAM runs →All figures at Q4_K_M quantization and 4K context. “Memory” is the model's download size from Hugging Face, plus its context cache calculated from the model's configuration, plus a runtime allowance of 10% of the model (at least 0.5GB, at most 3GB). How this is calculated
Image generators that also fit
These make pictures rather than text. Their memory use scales with the resolution you generate at, not conversation length, so treat these figures as a floor rather than a ceiling.
What you'd need more memory for
The next models up, and what they ask for.