Llama 4 Scout 17B-16E

Llama 4 Scout 17B-16E is a high-efficiency Mixture-of-Experts (MoE) model that features 17 billion active parameters out of 109 billion total parameters and 16 experts.

Llama 4 Scout 17B-16E is suited for retrieval tasks within long contexts and tasks that demand reasoning over large amounts of information, such as summarizing multiple large documents, analyzing extensive user interaction logs for personalization, and reasoning across large codebases.

Managed API (MaaS) specifications

Try in Agent Studio View model card in Model Garden

Model ID llama-4-scout-17b-16e-instruct-maas
Launch stage GA
Supported inputs & outputs
  • Inputs:
    Text, Code, Images
  • Outputs:
    Text
Capabilities
Usage types
Knowledge cutoff date August 2024
Versions
  • llama-4-scout-17b-16e-instruct-maas
    • Launch stage: GA
    • Release date: April 29, 2025
Supported regions

Model availability

  • United States
    • us-east5

ML processing

  • United States
    • Multi-region
Quota limits

us-east5:

  • Max output: 8,192
  • Context length: 1,310,720

Pricing See Pricing.

Deploy as a self-deployed model

To self-deploy the model, navigate to the Llama 4 Scout 17B-16E model card in the Model Garden console and click Deploy model. When you self-deploy from Model Garden, you can select from the following model variants:

  • llama-4-scout-17b-16e: Pretrained model.

  • llama-4-scout-17b-16e-instruct: Instruction-tuned model.

For more information about deploying and using partner models, see Deploy a partner model and make prediction requests.

Considerations

Keep the following considerations in mind when using Llama 4 Scout 17B-16E:

  • The managed MaaS endpoint (llama-4-scout-17b-16e-instruct-maas) serves the instruction-tuned model in fixed FP8 precision. You can't change the quantization format on the managed endpoint; to deploy a custom or pretrained variant, self-deploy the model from Model Garden.

  • You can include a maximum of three images per request.

  • The MaaS endpoint doesn't use Llama Guard, unlike earlier Llama versions. To use Llama Guard, deploy Llama Guard from Model Garden and then send the prompts and responses to that endpoint. Compared to Llama 4, Llama Guard has a smaller context window (128,000 tokens) and only accepts requests with a single image at the beginning of the prompt.