Llama 4 Scout debuts on Cloudflare Workers AI

Meta's Llama 4 family has arrived on Cloudflare's Workers AI platform, starting with the Llama 4 Scout model. As a launch partner, Cloudflare is making the open-source model available serverless, meaning developers can call it with a single API request without managing the underlying infrastructure.

Architecture: Mixture of Experts meets native multimodality

Llama 4 introduces a design combining a Mixture of Experts (MoE) architecture with an early-fusion backbone, making the model natively multimodal. Unlike its predecessor, Llama 3.2 11b, which used separate parameter sets for vision and text, Llama 4 parameters are trained to understand both modalities together. That training data includes large volumes of unlabeled text, image, and video content, so developers don't need to chain separate vision and language models to handle mixed input.

The Llama 4 family consists of two model sizes:

  • Llama 4 Scout: 109B total parameters, 17B active, 16 experts — available on Workers AI today
  • Llama 4 Maverick: 400B total parameters, 17B active, 128 experts

In an MoE model, a router component distributes input tokens to a subset of specialized neural networks, or "experts." For example, a request to write a Fibonacci algorithm in Haskell might be directed to an expert strong at code generation. The remaining experts stay unactivated, so Llama 4 Scout can handle that request using only one expert's 17B active parameters rather than the full 109B weight set. More complex queries will route to multiple experts, but the efficiency principle holds: MoE keeps inference speed high while preserving the quality gains of a much larger model.

A 10M-token context window

Llama 4 Scout's headline feature is a context window of up to 10,000,000 tokens, one of the first open-source models to reach that size. Such a window allows significantly longer conversations, more personalized outputs, and more effective Retrieval Augmented Generation (RAG) over large document sets or entire codebases. At launch, Workers AI supports a 131,000-token context window, with plans to expand further over time.

Hosting an MoE model poses challenges despite efficient training: serving requires loading the full set of model weights—over 200 GB—into GPU memory, and a larger context window requires additional Key Value cache memory. Workers AI abstracts those concerns away, handling hardware provisioning, memory, and infrastructure so developers can focus on the application layer.

Developers can experiment with Llama 4 Scout without creating an account via Cloudflare's AI playground, or consult the Workers AI developer docs for model details, pricing, and acceptable use terms.