Skip to content
AI Radar by Wagner Solution
CS
Menu
major New models verified update

Beam is available with open weights and through Amazon SageMaker JumpStart

confirmed by 2 independent sources updated 2 h ago

Reflection AI has released the weights for Beam under the Apache 2.0 license. The model for coding and agent tasks is also available through Amazon SageMaker JumpStart in two US regions. It offers a million-token context window, but self-hosting requires multiple GPUs.

Reflection AI has made the weights for Beam available on Hugging Face under the Apache 2.0 license. At launch, the model is also available through Amazon SageMaker JumpStart in the US East (N. Virginia) and US East (Ohio) regions. While the articles from The Decoder and TechCrunch still describe a release planned for later in the month, the current articles from Hugging Face and AWS already document the availability of the weights and deployment options.

Beam is designed for coding, reasoning and multistep agent tasks. It accepts and returns only text. Its mixture-of-experts architecture contains 501 billion parameters, of which 23 billion are activated for each token; the context window has a capacity of one million tokens. The model supports tool calling, preserving reasoning across turns and five reasoning levels: low, medium, high, xhigh and max. The release is accompanied by support from the transformers, SGLang and vLLM libraries, as well as an open tool for running agents written in Rust.

The weights are available in BF16, FP8 and NVFP4 formats. The BF16 files take up approximately 1 TB; during deployment, memory must accommodate all experts, even though only some are activated. According to AWS, the model requires multiple GPUs, and the BF16 variant requires planning for multiple nodes with headroom for the runtime. For FP8, AWS lists configurations with NVIDIA H100 or H200 GPUs; for NVFP4, it lists a configuration with NVIDIA B200 GPUs.

According to Reflection AI, Beam achieves results comparable to GLM-5.2 on demanding reasoning tests, with three to four times lower computational requirements during response generation. The TechCrunch article notes that these claims have not been independently verified.

What changed

Why it matters

Releasing the weights allows the model to be adapted and run in your own environment; availability through Amazon SageMaker JumpStart adds a managed deployment option. However, the low number of active parameters does not eliminate the memory requirements of the entire model, so looking only at computation per token is not enough when making infrastructure decisions.

Release card

Beam 1.0

Reflection AI

open weights
Specifications
The model has a total of 501 billion parameters, of which 23 billion are activated for each token.
Context
The context window has a capacity of 1 million tokens.
Inputs
The model accepts and returns only text.
Licence
Apache 2.0
Availability
The weights are publicly available on Hugging Face in BF16, FP8 and NVFP4 formats. The model can also be deployed through Amazon SageMaker JumpStart in the US East (N. Virginia) and US East (Ohio) regions. It is supported by the transformers, SGLang and vLLM libraries.
Documented measurements
  • SWEBench Verified 80.9 The results table published by Reflection AI lists a score of 80.9 for the model on this coding test.
  • Terminal Bench v2.1 80.1 The results table published by Reflection AI lists a score of 80.1 for the model on this agent task test.
  • SWE Bench Pro v1 65.5 The results table published by Reflection AI lists a score of 65.5 for the model on this coding test.
According to the sources, it is suitable for
  • Fixing software bugs and modifying code.
  • Multistep workflows with tool calling.
  • Adapting the model to custom tasks and deploying it on your own infrastructure.
Documented limits
  • Directly processing inputs other than text.
  • Deployment on a single GPU; the model requires multiple GPUs.

According to Reflection AI, the model achieves results comparable to GLM-5.2 on demanding reasoning tests, with three to four times lower computational requirements during response generation. The TechCrunch article notes that the performance claims made by the company have not been independently verified.

The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →

What was added since the original report

Verified updates

  1. New verified information

    Beam is available through Amazon SageMaker JumpStart at launch.; Availability in JumpStart includes the US East (N. Virginia) and US East (Ohio) regions.; BF16, FP8 and NVFP4 variants are available.

    • Beam is available through Amazon SageMaker JumpStart at launch.
    • Availability in JumpStart includes the US East (N. Virginia) and US East (Ohio) regions.
    • BF16, FP8 and NVFP4 variants are available.
  2. New verified information

    Beam has been released under the Apache 2.0 license.; Checkpoints are available in BF16, FP8 and NVFP4 formats.; The release is supported by the transformers, SGLang and vLLM libraries.; The model is accompanied by an open harness written in Rust.; The model supports five reasoning levels: low, medium, high, xhigh and max.

    • Beam has been released under the Apache 2.0 license.
    • Checkpoints are available in BF16, FP8 and NVFP4 formats.
    • The release is supported by the transformers, SGLang and vLLM libraries.
    • The model is accompanied by an open harness written in Rust.
    • The model supports five reasoning levels: low, medium, high, xhigh and max.
  3. New verified information

    Beam focuses on coding, logical reasoning and agent tasks.; According to the company, reinforcement learning ran for over four weeks on 10 500 Nvidia GB300 GPUs.; The release of the weights is planned for later in the month of the announcement.; The model is to be released under the Apache 2.0 license.

    • Beam focuses on coding, logical reasoning and agent tasks.
    • According to the company, reinforcement learning ran for over four weeks on 10 500 Nvidia GB300 GPUs.
    • The release of the weights is planned for later in the month of the announcement.
    • The model is to be released under the Apache 2.0 license.

Two audiences, two different impacts

What this means

01

For individuals

When solving a coding task, a developer can choose between five reasoning levels and incorporate tool calling into multistep work with code.

What to do Use a single coding task to compare the correctness of the result and response time at different reasoning levels.
More practical updates →
02

For a business

The team can choose between managed deployment and greater control over its own environment. Self-hosting requires infrastructure with multiple GPUs and capacity for all the weights as well as the runtime.

Development
What to decide Before deployment, check the compatibility of the selected weight variant, runtime and GPU, and allow for memory headroom beyond the size of the weights themselves.
More business impacts →
Amazon SageMaker HyperPod Amazon SageMaker JumpStart Apache 2.0 Beam mixture-of-experts MoE NVIDIA Reflection Reflection AI SGLang vLLM

Check the original

Event sources

confirmed by 2 independent sources · 4 publishers, 2 independent. We count feeds from the same owner only once.

6
TechCrunch AI independent context · first detected Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost The Decoder (daily AI news) independent context Reflection's Beam becomes the most capable open-weight model built outside China AI Business independent context Reflection’s Beam model signals deepening split in global AI market AI Business independent context Western open-weight campaign could give enterprises more control over AI An overview of multiple AI topics; AI Radar covers only this event. Hugging Face Blog primary source Welcome Beam: a 501B open model you can own, customize, and scale AWS Machine Learning Blog primary source Introducing Reflection Beam: an open-weight model for agentic coding on AWS