Skip to content

Best GPU for Popular AI Models: VRAM by Model

How much GPU memory Llama, Qwen, DeepSeek, Kimi, GLM, Mistral, Gemma, Phi, gpt-oss, FLUX.1, Stable Diffusion, Wan and Whisper need, and which GPUs hold them.

Each page computes the GPU memory of one open-weight model from its published architecture: weights in BF16, FP8 and 4-bit, KV cache at an 8,192-token context, LoRA, QLoRA and full fine-tuning. It then lists the GPUs that hold each setup on one to eight cards.

Llama (Meta)

Qwen (Alibaba Cloud)

  • Best GPU for Qwen3 4B

    4B parameters · Alibaba Cloud (Qwen team)

    BF16 10.2 GB · 4-bit 3.5 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Qwen2.5 7B

    7.6B parameters · Alibaba Cloud (Qwen team)

    BF16 17.3 GB · 4-bit 4.7 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Qwen3 8B

    8.2B parameters · Alibaba Cloud (Qwen team)

    BF16 19.4 GB · 4-bit 5.8 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Qwen2.5 VL 7B

    8.3B parameters · Alibaba Cloud (Qwen team)

    BF16 18.8 GB · 4-bit 5.1 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Qwen2.5 14B

    14.7B parameters · Alibaba Cloud (Qwen team)

    BF16 34.1 GB · 4-bit 9.9 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Qwen3 14B

    14.8B parameters · Alibaba Cloud (Qwen team)

    BF16 34 GB · 4-bit 9.6 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Qwen3 30B A3B

    30.5B parameters · Alibaba Cloud (Qwen team)

    BF16 68 GB · 4-bit 17.7 GB

    Inference: 1 GPU × NVIDIA A100 PCIe 40GB

  • Best GPU for Qwen3 Coder 30B A3B

    30.5B parameters · Alibaba Cloud (Qwen team)

    BF16 68 GB · 4-bit 17.7 GB

    Inference: 1 GPU × NVIDIA A100 PCIe 40GB

  • Best GPU for Qwen2.5 Coder 32B

    32.5B parameters · Alibaba Cloud (Qwen team)

    BF16 73.9 GB · 4-bit 20.2 GB

    Inference: 1 GPU × NVIDIA L40S 48GB

  • Best GPU for Qwen3 32B

    32.8B parameters · Alibaba Cloud (Qwen team)

    BF16 74.5 GB · 4-bit 20.4 GB

    Inference: 1 GPU × NVIDIA L40S 48GB

  • Best GPU for Qwen2.5 72B

    72.7B parameters · Alibaba Cloud (Qwen team)

    BF16 162.9 GB · 4-bit 42.9 GB

    Inference: 1 GPU × NVIDIA H200 141GB

  • Best GPU for Qwen2.5 VL 72B

    73.4B parameters · Alibaba Cloud (Qwen team)

    BF16 164.4 GB · 4-bit 43.3 GB

    Inference: 1 GPU × NVIDIA H200 141GB

  • Best GPU for Qwen3 235B A22B

    235B parameters · Alibaba Cloud (Qwen team)

    BF16 518.7 GB · 4-bit 131 GB

    Inference: 2 GPUs × NVIDIA B200 180GB

DeepSeek

Kimi (Moonshot AI)

  • Best GPU for Kimi K2

    1,026B parameters · Moonshot AI

    BF16 2,257.8 GB · 4-bit 564.9 GB

    Inference: 8 GPUs × NVIDIA B200 180GB

GLM (Z.ai)

  • Best GPU for GLM-4.5 Air

    106B parameters · Z.ai

    BF16 234.9 GB · 4-bit 60 GB

    Inference: 1 GPU × NVIDIA H200 141GB

  • Best GPU for GLM-4.5

    355B parameters · Z.ai

    BF16 784.4 GB · 4-bit 198.6 GB

    Inference: 2 GPUs × NVIDIA B300 288GB

Mistral AI

Gemma (Google)

  • Best GPU for Gemma 3 4B

    4.3B parameters · Google DeepMind

    BF16 10.7 GB · 4-bit 3.6 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Gemma 2 9B

    9.2B parameters · Google DeepMind

    BF16 23.4 GB · 4-bit 8.2 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Gemma 3 12B

    12.2B parameters · Google DeepMind

    BF16 30.4 GB · 4-bit 10.3 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Gemma 3 27B

    27.4B parameters · Google DeepMind

    BF16 64.9 GB · 4-bit 19.6 GB

    Inference: 1 GPU × NVIDIA A100 PCIe 40GB

Phi (Microsoft)

gpt-oss (OpenAI)

Image generation

Video generation

  • Best GPU for Wan 2.1 1.3B

    7.2B parameters · Alibaba (Wan team)

    BF16 15.9 GB · 4-bit 4 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for CogVideoX 5B

    10.5B parameters · Zhipu AI and Tsinghua University

    BF16 23.1 GB · 4-bit 5.8 GB

    Inference: 1 GPU × NVIDIA L4 24GB

  • Best GPU for Wan 2.1 14B

    20.1B parameters · Alibaba (Wan team)

    BF16 44.2 GB · 4-bit 11.1 GB

    Inference: 1 GPU × NVIDIA A100 PCIe 40GB

  • Best GPU for HunyuanVideo

    20.7B parameters · Tencent

    BF16 45.5 GB · 4-bit 11.4 GB

    Inference: 1 GPU × NVIDIA A100 PCIe 40GB

Speech recognition and embeddings

Explore the live market

Every offer on the live market comes from a reviewed supplier and is sold by EverXCloud Inc.