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 →

1792 GB/sGeForce RTX 5090 (32GB)· NVIDIA published specification: 1,792 GB/s GDDR7, 512-bit

Models that run on 32GB VRAM in 2026

ModelSizeMemoryEst. speedDownloadsFeels like
LFM2.5-230MLiquid Ai230M808 MBRuns well~7593 tok/s58,011Fast
MiniCPM5-1BOpenBMB1.1B1.3 GBRuns well~1785 tok/s667,154Fast
MiniCPM5-2BOpenBMB2.5B2.3 GBRuns well~746 tok/s206,774Fast
LFM2.5-2.6BLiquid Ai2.7B2.5 GBRuns well~696 tok/s98,713Fast
LFM2.5-VL-3BLiquid Ai3.1B2.5 GBRuns well~696 tok/s31,296Fast
North-Micro-Vision-InstructCohere2.5B2.5 GBRuns well~776 tok/s120,354Fast
granite-4.2-3bIbm3.7B3.1 GBRuns well~519 tok/s32,017Fast
Mage-VLMicrosoft4.7B3.9 GBRuns well~429 tok/s29,690Fast
LFM2.5-8B-A1BLiquid Ai · MoE8.5B5.9 GBRuns well~1913 tok/s1B active80,632Fast
Ling-3.0-tinyInclusionai · MoE7.9B6.2 GBRuns well~1207 tok/s1.6B active25,727Fast
granite-4.2-8bIbm8.8B6.6 GBRuns well~218 tok/s52,946Fast
Ornith-1.5-9BDeepreinforce9.7B6.9 GBRuns well~202 tok/s453,339Fast
gemma-4-12B-itGoogle DeepMind12B9.6 GBRuns well~161 tok/s2,958,440Fast
Qwen3.8-27BAlibaba / Qwen28B19 GBRuns well~71 tok/s7,703,400Fast
diffusiongemma-26B-A4B-itGoogle DeepMind · MoE26B19 GBRuns well~447 tok/s4B active708,393Fast
Muse-Glimmer-30BMeta AI30B20 GBRuns well~65 tok/s441,881Fast
granite-4.2-30bIbm29B21 GBRuns well~66 tok/s17,894Fast
North-Mini-Code-1.0Cohere · MoE30B22 GBRuns well~303 tok/s6.1B active11,742Fast
Laguna-XS-2.1Poolside · MoE33B23 GBRuns well~288 tok/s6.7B active32,445Fast
Nex-N2.5-miniNex Agi · MoE35B24 GBRuns well~275 tok/s7B active4,543Fast
Ornith-1.5-35B-A3BDeepreinforce · MoE36B24 GBRuns well~643 tok/s3B active359,808Fast
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4NVIDIA · MoE18B28 GBTight~275 tok/s3B active1,261,605Fast

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.

Other hardware

Verified against HuggingFace on 2026-09-15New to this? Start here →Runs on a laptop and up · last 120 days only