What can you run on a 32GB graphics card?

With 32GB VRAM you can run 65 of the 101 open-source models we track — up to about 36B parameters at Q4 quantization. The most popular that fits is Qwen3.5-9B, needing 6.8 GB and generating around ~205 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
granite-embedding-97m-multilingual-r2Ibm97M655 MBRuns well~11044 tok/s237,787Fast
granite-embedding-311m-multilingual-r2Ibm312M748 MBRuns well~6190 tok/s47,167Fast
LFM2.5-Encoder-230MLiquid Ai230M793 MBRuns well~8399 tok/s13,450Fast
LFM2.5-230MLiquid Ai230M808 MBRuns well~7593 tok/s54,986Fast
ERNIE-4.5-0.3B-PTBaidu AI361M853 MBRuns well~4840 tok/s21,323Fast
nemotron-3.5-asr-streaming-0.6bNVIDIA638M1.1 GBRuns well~2349 tok/s1,052,774Fast
Qwen3.5-0.8BAlibaba / Qwen873M1.3 GBRuns well~2187 tok/s2,990,867Fast
LFM2.5-1.2B-InstructLiquid Ai1.2B1.4 GBRuns well~1594 tok/s555,431Fast
LFM2.5-VL-1.6BLiquid Ai1.6B1.4 GBRuns well~1594 tok/s78,655Fast
Nemotron-3-Embed-1B-BF16NVIDIA1.1B1.5 GBRuns well~1691 tok/s467,453Fast
cohere-transcribe-arabic-07-2026Cohere2.1B1.9 GBRuns well~934 tok/s51,389Fast
Hy-MT2-1.8BTencent2B1.9 GBRuns well~1028 tok/s49,645Fast
Qwen3.5-2BAlibaba / Qwen2.3B2.0 GBRuns well~909 tok/s2,468,876Fast
granite-speech-4.1-2bIbm2.3B2.4 GBRuns well~747 tok/s402,320Fast
LFM2.5-2.6BLiquid Ai2.7B2.5 GBRuns well~696 tok/s77,973Fast
LFM2.5-Audio-1.5BLiquid Ai1.5B2.6 GBRuns well~639 tok/s1,025Fast
Unlimited-OCRBaidu AI3.3B2.7 GBRuns well~597 tok/s2,836,694Fast
LocateAnything-3BNVIDIA3.8B2.8 GBRuns well~553 tok/s463,853Fast
granite-4.1-3bIbm3.4B3.0 GBRuns well~555 tok/s300,243Fast
Shieldstral-1.0-3BMistral AI3.8B3.1 GBRuns well~543 tok/s2,480Fast
Ministral-3-3B-Instruct-2512Mistral AI3.8B3.1 GBRuns well~543 tok/s544,419Fast
Nemotron-Labs-Diffusion-3B-BaseNVIDIA3.8B3.3 GBRuns well~504 tok/s35,110Fast
Cosmos3-EdgeNVIDIA3.9B3.3 GBRuns well~500 tok/s53,138Fast
Nemotron-3.5-Content-SafetyNVIDIA4.3B3.6 GBRuns well~468 tok/s176,924Fast
gemma-4-E2B-itGoogle DeepMind5.1B3.8 GBRuns well~375 tok/s4,002,947Fast
Qwen3.5-4BAlibaba / Qwen4.7B3.8 GBRuns well~425 tok/s6,199,203Fast
Fara1.5-4BMicrosoft4.5B4.0 GBRuns well~404 tok/s4,882Fast
NVIDIA-Nemotron-3-Nano-4B-BF16NVIDIA4B4.1 GBRuns well~411 tok/s712,701Fast
Fara-7BMicrosoft8.3B5.5 GBRuns well~249 tok/s2,192Fast
Hy-MT2-7BTencent8B5.7 GBRuns well~252 tok/s26,629Fast
LFM2.5-8B-A1BLiquid Ai · MoE8.5B5.9 GBRuns well~1913 tok/s1B active170,464Fast
Nemotron-3-Embed-8B-BF16NVIDIA8B5.9 GBRuns well~243 tok/s142,381Fast
gemma-4-E4B-itGoogle DeepMind8B6.2 GBRuns well~218 tok/s5,295,850Fast
Nemotron-Labs-Diffusion-8B-BaseNVIDIA8.5B6.2 GBRuns well~227 tok/s413,039Fast
Ministral-3-8B-Instruct-2512Mistral AI8.9B6.3 GBRuns well~224 tok/s119,379Fast
granite-4.1-8bIbm8.8B6.3 GBRuns well~228 tok/s4,213,980Fast
Ornith-1.0-9BDeepreinforce9B6.7 GBRuns well~207 tok/s2,250,422Fast
Qwen3.5-9BAlibaba / Qwen9.7B6.8 GBRuns well~205 tok/s11,767,971Fast
GLM-4.6V-FlashZhipu AI10B7.0 GBRuns well~189 tok/s282,562Fast
Fara1.5-9BMicrosoft9.4B7.0 GBRuns well~197 tok/s14,526Fast
Olmo-3-7B-InstructAi27.3B7.2 GBRuns well~260 tok/s406,776Fast
gemma-4-12B-itGoogle DeepMind12B9.4 GBRuns well~164 tok/s2,963,990Fast
Ministral-3-14B-Instruct-2512Mistral AI14B9.7 GBRuns well~141 tok/s198,169Fast
Nemotron-Labs-Diffusion-14BNVIDIA14B9.7 GBRuns well~141 tok/s6,554Fast
Cosmos3-NanoNVIDIA16B11 GBRuns well~122 tok/s313,164Fast
ERNIE-4.5-21B-A3B-PTBaidu AI · MoE22B15 GBRuns well~631 tok/s3B active44,971Fast
Voxtral-Small-24B-2507Mistral AI24B16 GBRuns well~81 tok/s262,513Fast
Magistral-Small-2506Mistral AI24B16 GBRuns well~81 tok/s49,182Fast
Qwen3.5-27BAlibaba / Qwen28B19 GBRuns well~70 tok/s2,637,888Fast
diffusiongemma-26B-A4B-itGoogle DeepMind · MoE26B19 GBRuns well~447 tok/s4B active1,952,156Fast
gemma-4-26B-A4B-itGoogle DeepMind · MoE27B20 GBRuns well~456 tok/s4B active10,976,594Fast
ERNIE-4.5-VL-28B-A3B-PTBaidu AI · MoE29B20 GBRuns well~643 tok/s3B active218,879Fast
Qwen3.6-27BAlibaba / Qwen28B20 GBRuns well~68 tok/s6,572,759Fast
granite-4.1-30bIbm29B20 GBRuns well~67 tok/s20,228Fast
Fara1.5-27BMicrosoft27B20 GBRuns well~66 tok/s3,208Fast
Hy-MT2-30B-A3BTencent · MoE30B20 GBRuns well~640 tok/s3B active13,753Fast
North-Mini-Code-1.0Cohere · MoE30B22 GBRuns well~303 tok/s6.1B active23,720Fast
diffusiongemma-26B-A4B-it-NVFP4NVIDIA · MoE14B22 GBRuns well~222 tok/s4B active1,508,724Fast
GLM-4.7-FlashZhipu AI31B22 GBRuns well~62 tok/s2,071,716Fast
Olmo-3.1-32B-InstructAi232B23 GBRuns well~60 tok/s51,153Fast
Ornith-1.0-35BDeepreinforce · MoE35B24 GBRuns well~275 tok/s7B active2,736,644Fast
gemma-4-31B-itGoogle DeepMind31B24 GBRuns well~64 tok/s11,028,273Fast
Qwen3.5-35B-A3BAlibaba / Qwen · MoE36B25 GBRuns well~634 tok/s3B active2,539,444Fast
Qwen3.6-35B-A3BAlibaba / Qwen · MoE36B25 GBRuns well~631 tok/s3B active5,939,238Fast
Qwen-AgentWorld-35B-A3BAlibaba / Qwen · MoE35B25 GBRuns well~608 tok/s3B active84,349Fast

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.

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.

Other hardware

Verified against HuggingFace on 2026-08-08New to this? Start here →Runs on a laptop and up · last 12 months only