AWS released Reflection Beam, an open-weight model for agentic coding, with 501 billion total parameters, available in US East regions. It is the company's first major model release for agentic coding since its initial machine learning advancements.
AWS reported Reflection Beam has an effective context length of 1 million tokens, measured on long-horizon tasks such as maintaining code, using tools, and completing multi-step workflows. That compares with earlier models that had smaller context lengths.
Reflection Beam is built on AWS infrastructure and targets developers and researchers building software engineering and agentic systems. Availability begins with a public preview, initially for developers and researchers using Amazon SageMaker JumpStart.
"Beam brings this approach to developers building software engineering and agentic systems," said Reflection's mission statement. The model is released through Hugging Face under the Apache 2.0 license.
The announcement follows AWS's expansion of its machine learning offerings. AWS said the model is part of its broader strategy to make intelligence open and accessible.
AWS did not say when the model will be available in other regions, and it raised the question of how developers will manage the model's large parameter count. The model will be available in BF16, FP8, and NVFP4 checkpoint variants.
Source: awsml
