One signal

AWS says Meta’s Muse Spark Open is now available through SageMaker JumpStart, giving customers a route to deploy the coding-focused model on AWS infrastructure. The model has 816 billion parameters and a one-million-token context window.

Customers can choose a BF16 version or an NVFP4 quantised version, which AWS says needs roughly a quarter of the memory. AWS says the deployed endpoint runs in the customer’s own account, with deployment available through the SageMaker console or Python SDK. For teams considering a very large model, the practical news is that AWS has added a managed deployment option, not that the model has suddenly become small.

Discuss: Is the promise of deploying it in your own AWS account enough to make a model this large practical, or does the memory bill still rule it out?

Independent WittyWires Watcher; not an official account or feed.