HomePricingBlogsDocs

Open-weight AI models you can run on a GPU pod

VRAM requirements and the cheapest live GPU fit for 1,429 Hugging Face models, grouped by model line and publisher, at native, FP8 and INT4 precision. Aquanode doesn't run a hosted inference API: every model page links to a GPU pod sized for it, billed per second, that you install your own inference stack on (or, for image and video models, our ComfyUI template). Not sure which model to run in the first place? See how the leading open-weight families score on published benchmarks in the LLM benchmarks leaderboard.

How the VRAM and cost numbers are calculated

  • Weight bytes = parameter count × bytes-per-parameter for the shown precision (FP32=4, BF16/FP16=2, FP8=1, INT4=0.5 bytes).
  • Required VRAM adds a flat 1.2× overhead on top of weight size for activations and allocator fragmentation, plus KV-cache for text and text+image (VLM) models, calibrated for an assumed 8,192-token context. It's a rule-of-thumb, not a per-request simulation: long contexts or large batch sizes need materially more headroom for language models, diffusion/video models scale mainly with output resolution and frame count instead of context length, and speech models scale mainly with input audio length.
  • “Cheapest live fit” only matches a GPU to a row if its hardware actually supports that row's compute format, and caps a fit at 8 GPUs of the same model. A pre-Ampere card (Pascal/Volta/Turing) never qualifies for a BF16 or FP8 row, regardless of VRAM or price. See the GPU Index for per-model specs.
  • The INT4 row additionally requires a quantized checkpoint that's actually been published for the model in question. For language models this also means a GPU with AWQ/GPTQ/Marlin-class quantized-serving kernel support. Neither is guaranteed, so check the model's Hugging Face page for a published quantized variant before relying on that row.
  • Prices are live per-GPU $/hr rates from Aquanode's live price feed, normalized the same way as the GPU Index, with outlier offers excluded.
  • Multi-GPU fits assume the model shards cleanly across cards (e.g. tensor parallelism). Real throughput and interconnect overhead aren't modeled.

Qwen

Qwen3

104 models · 537M to 480.2B

Qwen2.5

56 models · 135M to 73.4B

Qwen3.5

46 models · 556M to 403.4B

Qwen3.6

16 models · 386M to 36.0B

Qwen3.8

38 models · 1.9B to 2446.2B

Qwen2

15 models · 494M to 72.7B

Qwen

4 models · 782M to 72.3B

Qwen-Image

7 models · 338M to 27.4B

Qwen1.5

18 models · 464M to 72.3B

Meta

Llama 3.2

21 models · 506M to 88.6B

Llama 3.1

42 models · 950M to 405.9B

Llama 3

22 models · 8.0B to 70.6B

Llama 2

11 models · 6.7B to 13.0B

Llama 3.3

4 models · 49.9B to 70.6B

Code Llama

3 models · 6.7B to 33.7B

Llama

3 models · 134M to 6.7B

Llama Guard

4 models · 1.2B to 12.0B

Google

Gemma 4

42 models · 420M to 32.7B

Gemma 3

13 models · 268M to 27.4B

Gemma 2

16 models · 2.5B to 27.2B

MedGemma

4 models · 4.3B to 28.8B

PaliGemma

3 models · 2.9B to 2.9B

Gemma 3n

4 models · 5.4B to 8.4B

Gemma

3 models · 2.5B to 8.5B

Stability AI

Stable Diffusion XL

33 models · 611M to 2.6B

Stable Diffusion

9 models · 860M to 1.3B

Stable Diffusion 3

3 models · 2.1B to 8.1B

StableLM

3 models · 1.6B to 2.8B

stabilityai

3 models · 568M to 1.3B

NVIDIA

Parakeet

7 models · 609M to 1.1B

Nemotron 3

16 models · 638M to 560.5B

Cosmos

6 models · 2.4B to 64.6B

Nemotron Nano

4 models · 8.9B to 13.2B

Nemotron

5 models · 618M to 31.6B

Nemotron Labs

4 models · 3.8B to 13.5B

Nemotron-H

3 models · 8.1B to 56.3B

Z.ai

GLM-5

22 models · 1.2B to 753.9B

GLM-4.5

12 models · 10.3B to 358.5B

GLM-4

6 models · 9.4B to 32.6B

CogVideoX

3 models · 1.7B to 5.6B

DeepSeek

DeepSeek R1

13 models · 1.8B to 684.5B

DeepSeek V3

8 models · 684.5B to 685.4B

DeepSeek Coder

6 models · 1.3B to 15.7B

DeepSeek V2

5 models · 15.7B to 235.7B

DeepSeek VL

4 models · 2.0B to 7.3B

Meta

Wav2Vec 2.0

32 models · 315M to 963M

facebook

3 models · 303M to 2.3B

MMS

3 models · 315M to 965M

IBM

Granite 4

21 models · 352M to 32.2B

Granite 3

12 models · 1.3B to 8.6B

Granite

4 models · 258M to 3.5B

Microsoft

Florence-2

5 models · 232M to 777M

Phi-4

7 models · 3.8B to 14.7B

Phi-3.5

4 models · 3.8B to 41.9B

Phi-3

6 models · 3.8B to 14.0B

VibeVoice

4 models · 323M to 8.7B

microsoft

6 models · 176M to 4.7B

OpenAI

Whisper

30 models · 242M to 1.6B

gpt-oss

6 models · 350M to 116.8B

Ai2

OLMo 2

5 models · 1.4B to 32.2B

Olmo 3

8 models · 7.3B to 32.2B

OLMoE

4 models · 6.9B to 6.9B

Molmo

3 models · 4.9B to 8.7B

EleutherAI

Pythia

15 models · 213M to 12.0B

GPT-Neo

3 models · 150M to 2.7B

Hugging Face

SmolLM2

6 models · 135M to 1.7B

SmolVLM

4 models · 256M to 2.2B

SmolLM3

3 models · 3.1B to 3.1B

SmolLM

5 models · 135M to 1.7B

ornith-ai

Ornith 1

18 models · 4.3B to 403.4B

Wan-AI

Wan2.1

11 models · 1.4B to 17.3B

Wan2.2

7 models · 5.0B to 16.3B

Liquid AI

LFM2

16 models · 230M to 8.5B

Mistral AI

Mistral 7B

8 models · 7.2B to 7.2B

Mistral Small

4 models · 23.6B to 24.0B

TII

Falcon

4 models · 6.9B to 41.8B

Falcon-H1

4 models · 521M to 7.6B

Falcon3

4 models · 1.7B to 7.5B

Moonshot AI

Kimi K3

3 models · 2.2B to 3.6B

Kimi K2

5 models · 3.0B to 1026.9B

OpenBMB

MiniCPM-V

3 models · 1.3B to 8.7B

MiniCPM5

3 models · 1.1B to 1.1B

MiniCPM4

4 models · 434M to 8.2B

Tencent

Hy3

3 models · 298.8B to 298.8B

Hy-MT2

3 models · 2.0B to 30.1B

01.AI

Yi-1.5

6 models · 6.1B to 34.4B

Yi

5 models · 6.1B to 34.4B

MiniMax

MiniMax M2

4 models · 228.7B to 228.7B

MiniMax M3

4 models · 3.1B to 440.3B

Black Forest Labs

FLUX.1

5 models · 347M to 11.9B

FLUX.2

5 models · 3.9B to 32.2B

inclusionAI

LLaDA

5 models · 8.0B to 16.3B

Ling

3 models · 7.9B to 127.5B

llava-hf

LLaVA

9 models · 894M to 13.4B

openai-community

GPT-2

8 models · 110M to 1.6B

LG AI Research

EXAONE 3.5

4 models · 2.4B to 32.0B

EXAONE 4.0

3 models · 1.3B to 32.0B

OpenGVLab

InternVL2

4 models · 938M to 25.5B

InternVL3

4 models · 938M to 78.4B

internlm

InternLM2

5 models · 7.7B to 19.9B

Nous Research

Hermes 3

3 models · 8.0B to 405.9B

Hermes 2

4 models · 6.7B to 46.7B

ATH-MaaS

Ovis

6 models · 853M to 10.2B

bigscience

BLOOM

6 models · 559M to 176.2B

Cohere

CohereLabs

3 models · 2.1B to 30.5B

typhoon-ai

Typhoon

6 models · 2.1B to 30.5B

upstage

SOLAR

6 models · 10.7B to 250.3B

Xiaomi MiMo

MiMo V2

4 models · 9.4B to 1023.2B

Lightricks

LTX

5 models · 1.9B to 19.0B

Salesforce

BLIP

3 models · 470M to 3.9B

arcee-ai

Trinity

3 models · 6.1B to 398.6B

Baidu

ERNIE 4.5

3 models · 361M to 21.9B

KBlueLeaf

KBlueLeaf

4 models · 393M to 508M

parler-tts

Parler-TTS

4 models · 878M to 2.3B

poolside

Laguna

4 models · 33.4B to 117.6B

StepFun

stepfun-ai

4 models · 561M to 321.0B

Zyphra

Zamba

4 models · 1.2B to 7.4B

ai21labs

Jamba

3 models · 3.0B to 51.6B

ByteDance

ByteDance

3 models · 1.4B to 2.7B

ByteDance Seed

ByteDance-Seed

3 models · 8.3B to 36.2B

convaiinnovations

convaiinnovations

3 models · 322M to 421M

datalab-to

datalab-to

3 models · 686M to 9.7B

fastino

GLiNER

3 models · 287M to 486M

FastVideo

FastVideo

3 models · 5.0B to 35.0B

google-t5

T5

3 models · 223M to 2.9B

h2oai

h2oai

3 models · 514M to 2.2B

HuggingFaceM4

Idefics

3 models · 8.4B to 8.9B

lightonai

lightonai

3 models · 1.0B to 1.2B

meituan-longcat

LongCat

3 models · 69.1B to 561.9B

Nanbeige

Nanbeige

3 models · 848M to 4.2B

naver-hyperclovax

naver-hyperclovax

3 models · 3.7B to 33.3B

skt

skt

3 models · 7.3B to 691.7B

state-spaces

Mamba

3 models · 129M to 2.8B

swiss-ai

Apertus

3 models · 8.1B to 70.6B

utter-project

EuroLLM

3 models · 1.7B to 9.2B

vikhyatk

Moondream

3 models · 1.9B to 9.3B

More models

Single models and small groups of models from one publisher.

VRAM calculatorFine-tune a modelPrototype in a notebookBrowse the GPU Index

Submit the job. Everything after that is ours.

Sign up in 60 seconds. Pay for the GPU minutes you actually use.

Aquanode LogoAquanode

Run AI without spending any time on infrastructure.

© 2026 Aquanode. All rights reserved.

All trademarks, logos and brand names are the property of their respective owners.