On October 16, Reflection announced that it has released the weights
of its AI model Beam under the Apache 2.0 license.
Developers can download the model and run it themselves,
fine-tune it for their work, and connect it to AI agents.
The model is also available through Reflection’s API
and major cloud services, among other options.
What Changes: From Introducing the Model to Making the Weights Available
Beam is a model built for coding, reasoning,
and AI agent use cases that carry out multi-step work.
Reflection introduced the model on October 5,
and has now released the weights and related materials,
moving to a stage where developers can run, evaluate, and customize it themselves.
Along with the model card, the company has also published
the Beam Safety Technical Report,
which summarizes its safety evaluation methods and results.
Beam has about 501 billion total parameters
and uses a Mixture-of-Experts (MoE) architecture
that selectively activates about 23 billion of them during processing.
According to the model card, its context length is 1 million tokens.
Separate Paths for Self-Hosting and API Use
The weights are published on Hugging Face
in BF16, FP8, and NVFP4 numerical formats.
The model also supports inference software such as vLLM and SGLang.
Reflection’s own API is offered as a public beta,
and users can start for free by signing up on the company’s platform.
According to the announcement, Beam launched with more than 40 partners.
Along with AWS, Google Cloud, and Microsoft Azure,
the company lists CoreWeave, Databricks, OpenRouter, and others.
Deployment in on-premises and other environments
is also supported through Dell and HPE.
Reflection also introduced Mirror CLI,
a tool for carrying out development work from the terminal.
Developers can inspect what the agent is doing, approve tool calls,
and steer the agent while it works.
The company has also tested Beam with other agent tools,
including OpenCode and Hermes.
Scope of Use: Commercial Use and Model Customization in Mind
The model card says Beam is intended for commercial and research use,
and lists fine-tuning, quantization,
and using its outputs to train other models,
among other activities, as permitted.
It also states that the model handles text only
and was developed predominantly for English.
LC Signal’s Perspective: Weights Together with the Tools to Run Them
Beyond the model weights, this release covered support for inference software,
fine-tuning, and connections to agents all at once.
Offering both self-hosted customization and use through an external API
seems likely to give organizations options
as they move from trials to building the model into their work.
Key Sources
-
Reflection official announcement, October 16, 2026:
Beam is now available
-
Reflection official model card:
Beam Model Card
-
Reflection official announcement, October 5, 2026 (initial introduction):
Introducing Beam: Reflection’s 501B open-weight model
-
Reflection official safety technical report:
Beam Safety Technical Report
This article was prepared with AI assistance using publicly available information
as of October 17, 2026.
Announcement dates follow the dates shown on the official pages.
This news covers the October 16 release of the weights and launch of availability,
and is distinguished from the October 5 initial introduction.
The release of the model weights and the public beta of the API
are described as separate forms of availability.
“LC Signal’s Perspective” reflects this publication’s interpretation
of the available information.