Beam is planned for release under the Apache 2.0 license, with an early version available to selected users
Reflection plans to release the weights for Beam under the Apache 2.0 license later this month. An early version is available to selected users, and final safety testing continues. The model is intended for coding, reasoning and agent tasks.
Reflection plans to release the weights for Beam under the Apache 2.0 license later this month, along with a technical report and developer documentation. The model is undergoing final safety testing; an early version is currently available to selected users. According to the company, the reinforcement learning phase lasted more than four weeks and used 10 500 Nvidia GB300 GPUs.
Beam is a text model for coding, logical reasoning and agent tasks. It uses a mixture-of-experts architecture: out of a total of 501 billion parameters, it activates 23 billion per token. The context window holds one million tokens, and pretraining used 23.8 trillion tokens, according to the company.
According to Reflection, Beam achieves results comparable to GLM-5.2 on demanding reasoning tasks with three to four times lower compute requirements during inference. These claims have not been independently verified. Users can set the reasoning depth to choose between response speed, compute costs and output quality.
Why it matters
The planned release of the weights matters for teams that want to run and customize the model on their own infrastructure. However, deployment decisions will depend on verifying quality and costs on their own tasks; the claim of lower compute requirements alone does not yet demonstrate actual operating savings.
Release card
Beam
Reflection
- Specifications
- The model has a total of 501 billion parameters and activates 23 billion parameters per token.
- Context
- The context window has a capacity of 1 million tokens.
- Inputs
- The model processes text only. According to Reflection, it can process content from other media after that content has been converted to text.
- Licence
- Pro plánované zveřejnění vah je uvedena licence Apache 2.0.
- Availability
- An early version is available to selected users, and final safety testing is underway. Reflection plans to release the weights, a technical report and developer documentation later in the month of the announcement.
- DeepSWE v1.1 44.4 Reflection reports a score of 44.4 for Beam on this coding test using agents; the results have not been independently verified.
- SWE Bench Pro v2-Hard 77.2 Reflection reports a score of 77.2 for Beam on this coding test using agents; the results have not been independently verified.
- SWE Bench Pro v1 65.5 Reflection reports a score of 65.5 for Beam on this coding test using agents; the results have not been independently verified.
- Terminal Bench v2.1 80.1 Reflection reports a score of 80.1 for Beam on this coding test using agents; the results have not been independently verified.
- SWE Atlas Codebase QnA 34.6 Reflection reports a score of 34.6 for Beam on this coding test using agents; the results have not been independently verified.
- Coding and software engineering tasks.
- Solving logical reasoning tasks.
- Automated workflows using agents.
- Directly processing other media without converting their content to text.
Reflection positions Beam against models from DeepSeek and Qwen, with an emphasis on computational efficiency. According to the company, it achieves results comparable to GLM-5.2 on demanding reasoning tasks with three to four times lower compute requirements during inference. These claims 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
-
Beam focuses on coding, logical reasoning and agent tasks.; According to the company, reinforcement learning ran for more than 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 planned for release under the Apache 2.0 license.
- Beam focuses on coding, logical reasoning and agent tasks.
- According to the company, reinforcement learning ran for more than 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 planned for release under the Apache 2.0 license.
Two audiences, two different impacts
What this means
For individuals
A developer with access to the early version can adjust reasoning depth for coding tasks depending on whether they need a quick answer or a more thorough solution.
For a business
For companies automating workflows, the planned open weights and Apache 2.0 license offer the possibility of self-hosting. The claimed reduction in compute requirements cannot yet be treated as a verified basis for an operating budget.
ProcessesCheck the original
Event sources
confirmed by 2 independent sources · 2 publishers, 2 independent. We count feeds from the same owner only once.