Skip to navigation
Model Recipes

Model Recipes

View as Markdown

Model Recipes

Browse ready-to-use recipes for tuning, validating, benchmarking, and deploying models. Choose a model below to find configurations and commands you can adapt with your own compatible checkpoint and data.

Find a model

TaskModelHubPrecisionOffloadingSingle GPUMulti-GPUDeployment
Image classificationResNet-50timmFP16-Ahead of time, Runtime-Dynamo + Triton
Image generationFLUX.1-devHugging FaceBF16, FP8 and NVFP4-RuntimeRuntimeDynamo
Stable Diffusion 3 MediumHugging FaceFP16-Ahead of time, Runtime-Dynamo
Stable Diffusion XLHugging Face--Ahead of time, Runtime--
Object detectionYOLOv10nHugging Face--Ahead of time-Triton
Protein language modelingESM-2 650MHugging Face--Ahead of time, Runtime-Dynamo
Speech recognitionParakeet CTC 0.6BHugging Face--Ahead of time, Runtime-Dynamo
Parakeet RNNT 1.1BHugging Face--Ahead of time, Runtime-Dynamo
Text encodingBERT-base-uncasedHugging Face--Ahead of time-Triton
Text embeddingsE5 Large V2Hugging Face--Ahead of time-Dynamo
Text generationQwen3 0.6BHugging Face--RuntimeEager only-
Video generationWan 2.1 T2V 1.3BHugging FaceBF16; FP32 VAE-Ahead of time, RuntimeAhead of time, RuntimeDynamo
  • Model: links to the recipe and usage instructions.
  • Hub: links to the source model card or collection.
  • Precision: numerical formats available in the recipe, such as FP16, BF16, FP8, and NVFP4.
  • Offloading: whether the recipe supports moving model weights between GPU and CPU memory.
  • Single GPU / Multi-GPU: compilation options for each setup. Ahead of time uses prebuilt backend artifacts; Runtime compiles during loading or inference; Eager only runs without a compiled backend.
  • Deployment: documented serving targets, such as Triton and Dynamo.
  • -: no option is documented in the current recipe.

Compilation options describe the main tuned component. Open the model page for supported combinations, hardware requirements, and results.