🔍 Read the full analysis: How To Own, Customize, And Scale The Welcome Beam 501B Open Model on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Reflection has released Beam, a text-only, 501-billion-parameter mixture-of-experts model under the Apache 2.0 license. The company describes it as a model for coding and agentic tasks, with 23 billion parameters active per token and a context window of up to 1 million tokens; independent performance results and a complete hardware-cost breakdown were not provided in the announcement.
Reflection has released Beam, a text-only model with 501 billion total parameters, on Hugging Face under the Apache 2.0 license. The company says Beam is intended for coding and agentic tasks and supports a stated context window of up to 1 million tokens. The release also includes integrations and adaptation recipes, as described in the original analysis, but it does not provide independent benchmark results to verify the model’s performance.
Beam uses a mixture-of-experts (MoE) architecture: 23 billion parameters are active for each token, while the full model has 501 billion. According to the release materials, its 52 layers include 256 experts in each MoE layer, with eight selected for a given token. Sparse activation reduces how many parameters are used in an individual computation, but the release notes that all experts still need to be kept in memory for deployment.
The BF16 checkpoint is about 1 terabyte in safetensors format. Reflection lists BF16, FP8 and NVFP4 checkpoint formats, saying NVFP4 is intended for NVIDIA Blackwell GPUs and FP8 versions are available for Hopper and Ampere chips. The announcement does not give a complete minimum-hardware specification or quantify the quality and speed trade-offs among those formats.
Beam has support at release in Transformers, vLLM and SGLang, as well as Reflection’s Rust-based harness. The materials also describe recipes for inference, supervised fine-tuning and reinforcement learning, including full-parameter reinforcement learning with Miles. Beam takes and produces text, offers tool calling with a dedicated grammar, and includes reasoning settings from low to max. A feature called “preserve_thinking” is described as carrying reasoning across turns.
What Beam Offers Model Builders
Apache 2.0 licensing gives developers permission to use, modify and redistribute Beam subject to the license’s terms. That can give organizations more choice over adapting a model for internal workflows and running it on their own infrastructure or through a cloud provider. The published integrations and training recipes may make it easier for teams with suitable computing resources to try those options.
But the distinction between 23 billion active parameters and 501 billion total parameters matters for planning. Active parameters describe part of the computation for each token, not the full memory footprint. The roughly 1 TB BF16 checkpoint signals a substantial hosting requirement, while the announcement leaves deployment costs and the performance effects of lower-precision formats unquantified.
Reflection and Hugging Face have also announced $10,000 in Hub credits for selected projects using Beam. The listed eligible work includes demos, quantizations, datasets, reinforcement-learning environments and benchmarks shared on Hugging Face. These projects may provide examples of use, but vendor recipes or community demonstrations alone cannot establish comparative model quality.
high performance GPU server for AI models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Beam Handles Long Inputs
The release describes a hybrid attention design: 43 layers attend to a sliding window of 512 prior tokens, while nine layers can attend across the full context. Reflection presents this design alongside the claimed context length of up to 1 million tokens. That figure describes a stated capability; the supplied materials do not include independent tests of accuracy, speed or cost at that length.
Within the MoE layers, a router selects eight of 256 experts for each token. The release says those experts account for 98.4% of the model’s parameters. The design uses a large pool of experts while limiting which ones are involved in a particular token’s computation; it does not remove the need to store the full set of weights when serving the model.
The release materials describe Beam as Reflection’s first model and identify Apache 2.0 as its license. Hugging Face says it worked with Reflection on model support and recipes for running and adapting Beam. The announcement establishes the available tools and stated architecture, but it is not an independent evaluation of the model.
“501B-parameter MoEs LLM with 23B active per token.”
— Reflection and Hugging Face, in the Beam release materials
large capacity NVFP4 GPU memory modules
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and Hardware Costs
The announcement does not include independent benchmark results, comparisons with other models or task-level evidence for coding and agentic work. It also provides no reported quality, latency or memory measurements for using the stated 1-million-token context window. Those missing results mean readers cannot assess Beam’s relative performance from the release alone.
Practical hardware needs will vary with the checkpoint format, serving configuration and workload. Although the materials give an approximate BF16 checkpoint size and name GPU generations for FP8 and NVFP4 versions, they do not set out a minimum hardware configuration or deployment cost. They also do not quantify how lower precision affects output quality or speed. The scale and timing of any independent testing are not specified.
As an affiliate, we earn on qualifying purchases.
Testing and Hub Credit Awards
Developers can use the published library integrations and recipes to run Beam or try post-training. Reflection and Hugging Face say they will collect links to qualifying projects shared on the model’s Hub page and select projects to receive credits in their Hub accounts. The supplied announcement does not specify when submissions close or when awards will be decided.
The next useful evidence would come from evaluations that report tasks, prompts, hardware, checkpoint format and measurement methods. Until those details are available, Beam’s listed capabilities and architecture are confirmed as release claims, while its comparative performance and real-world deployment costs remain open questions.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Beam?
Beam is Reflection’s text-only, 501-billion-parameter mixture-of-experts model, released on Hugging Face under the Apache 2.0 license. Reflection says it is designed for coding and agentic tasks.
Does Beam need 501 billion parameters for every token?
No. The release says 23 billion parameters are active per token, with eight experts selected from 256 in each MoE layer. However, the full set of experts must still be held in memory for deployment, according to the materials.
What hardware does Beam require?
The announcement gives a BF16 checkpoint size of about 1 terabyte and lists FP8 and NVFP4 formats for specified GPU generations. It does not provide a complete minimum-hardware specification or cost estimate.
Has Beam’s performance been independently verified?
The supplied release materials do not include independent benchmark results or comparisons with other models. The stated coding, agentic and long-context capabilities should be distinguished from independently tested results.
How can developers try Beam?
The model is available on Hugging Face, with support listed for Transformers, vLLM and SGLang, plus Reflection’s Rust-based harness. The release also describes inference and post-training recipes.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
