Reflection Beam: open coding model arrives on AWS

Reflection AI's coding model, Beam, is now available on Amazon's cloud. Its weights, the files that make up the model, can be downloaded for free under the Apache 2.0 license. For developers, that means a large open model they can run in their own cloud account.

What happened

Amazon Web Services (AWS) published a guide on 16 October 2026 to Beam, an open-weight model for "agentic coding." That means jobs where an AI keeps working on its own: it maintains code, calls tools, reads the results and finishes multi-step tasks.

AWS says the weights are on Hugging Face under the Apache 2.0 license. At launch, Beam is available in Amazon SageMaker JumpStart in two US regions, US East (N. Virginia) and US East (Ohio).

The guide describes three ways to run it. One is a managed JumpStart endpoint. Another is JumpStart on HyperPod Inference. The third uses SGLang or vLLM on HyperPod for more control over the setup.

Key facts

• Size: 501 billion parameters in total, with 23 billion active for each token, in a sparse "mixture of experts" design (per AWS).
• Context: Reflection reports an effective context of 1 million tokens. That is the company's own figure.
• License: Apache 2.0, with weights on Hugging Face, according to AWS.
• Versions: BF16, FP8 and NVFP4. They use different amounts of memory.
• Hardware: BF16 needs more than one machine. FP8 runs on NVIDIA H100 or H200 machines. NVFP4 runs on NVIDIA B200 machines.
• AWS gives no price and no benchmark scores in this guide.

Background

Reflection AI announced Beam on 5 October 2026, per DataCamp. It called Beam its first open-weight model, with weights promised under Apache 2.0 later in October. As of 6 October, DataCamp says, they were not yet released. The AWS guide, ten days later, says they are now on Hugging Face.

DataCamp also reports Reflection's own numbers. These come from the company, and nobody had checked them independently at the time. Beam scored 80.9 on SWE-bench Verified, a coding test. Reflection says it matches another open model on advanced reasoning while using three to four times less computing power. DataCamp adds that Beam trails some rivals on harder agentic tasks.

DataCamp also says the context the model was trained at was 256K tokens, and its beta API limits input plus output to 262,144 tokens. So the 1 million figure may not be what you get in every setup.

What it means for you

If you do not write code, nothing changes for you today. If you do, you now have a big coding model with a permissive license that you can run yourself instead of renting from a closed company.

Our take: it is not a plug-and-play tool. The guide says the largest versions need top-end NVIDIA machines, a GPU quota in your AWS account, and Python 3.11 or newer. It also warns you to delete endpoints and cached files after testing to avoid ongoing charges.

FAQ

What is Reflection Beam?

It is Reflection AI's first open-weight model, built for coding agents. It has 501 billion parameters in total and 23 billion active at a time.

Can I download Beam for free?

AWS says the weights are on Hugging Face under the Apache 2.0 license. Running it on AWS costs money, though, and the guide gives no prices.

Where can I run it on AWS?

At launch it is in Amazon SageMaker JumpStart in US East (N. Virginia) and US East (Ohio), and on HyperPod Inference.

Sources

• AWS ML Blog
• datacamp.com