🔍 Read the full analysis: Welcome Beam: A 501B Open Model You Can Own, Customize, And Scale on ThorstenMeyerAI.com
TL;DR
Reflection has released Beam, a text-only, 501-billion-parameter mixture-of-experts model under the Apache 2.0 license. The company says it targets coding and agentic tasks and supports up to a 1-million-token context, but the supplied release materials include no independent benchmarks and describe a BF16 checkpoint of about 1 terabyte.
Reflection has released Beam, a text-only, 501-billion-parameter mixture-of-experts model, as the original analysis describes, on Hugging Face under the Apache 2.0 license. The company says the model is designed for coding and agentic tasks and supports a context window of up to 1 million tokens; the release materials do not include independent performance results.
Beam has 23 billion parameters active per token, according to the release materials. Its 52 layers include 256 experts in each mixture-of-experts layer, with eight selected for each token. This sparse design limits which experts are used in an individual computation, but does not eliminate the need to keep the full set of model weights available for deployment.
The BF16 checkpoint is about 1 terabyte in safetensors format. Reflection lists BF16, FP8 and NVFP4 formats, with NVFP4 intended for NVIDIA Blackwell GPUs and FP8 versions offered for Hopper and Ampere chips. The announcement does not provide a full minimum-hardware or operating-cost breakdown, or quantify performance and quality differences between formats.
Beam has support from launch in Transformers, vLLM and SGLang, as well as Reflection’s Rust-based harness. The release also describes recipes for inference, supervised fine-tuning and reinforcement learning, including full-parameter reinforcement learning with Miles. The model accepts and produces text, and its listed features include tool calling and adjustable reasoning settings.
What Beam Opens to Developers
The Apache 2.0 license allows use, modification and redistribution subject to its terms, giving developers and organizations options to adapt Beam and choose where to run it. The inference and post-training recipes, along with support for common libraries, provide routes for teams to experiment with the model rather than relying only on a hosted service.
Those options come with substantial infrastructure demands. The distinction between 23 billion active parameters and 501 billion total parameters matters: sparse computation can reduce the work done for each token, while the full weights still require memory. The roughly 1 TB BF16 checkpoint makes clear that having an open license does not by itself make the model practical to host for every team.
Reflection and Hugging Face also announced $10,000 in Hub credits for selected projects using Beam. Eligible work may include demos, quantizations, datasets, reinforcement-learning environments and benchmarks. Projects could offer examples of use, but submissions and vendor-provided recipes would not substitute for comparable, independently documented evaluations.
high-performance GPU for AI training
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 says Beam supports a context window of up to 1 million tokens. That figure is a stated capability; the materials provided do not report independent tests of accuracy, latency or memory use at that length.
Within the mixture-of-experts layers, a router selects eight of 256 experts for each token. Reflection’s materials say those experts account for 98.4% of the model’s parameters. The architecture is intended to make use of a large pool of learned capacity without activating every expert for every token, but the full model’s storage requirements remain relevant to serving it.
The announcement presents Beam as Reflection’s first model and describes its release as open. Hugging Face worked with Reflection on support for running and adapting it. The listed license is Apache 2.0, and the release includes materials for inference and training workflows; the supplied source does not establish comparative results against other models.
“501B-parameter MoEs LLM with 23B active per token.”
— Reflection and Hugging Face, in the Beam release materials
large language model inference server
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and Hardware Remain Open
The supplied announcement does not include independent benchmark results, comparisons with other models, or measured performance on coding and agentic tasks. It also does not provide quality, speed or memory measurements for use with long inputs, including at the stated 1-million-token context length. The model’s published features and architecture do not establish how it performs in practical workloads.
Hardware requirements will vary with checkpoint format, serving setup and workload. Although the release gives a BF16 file size of about 1 terabyte and identifies GPU generations for some lower-precision formats, it does not specify a complete minimum configuration, deployment cost or the quality trade-offs of using FP8 or NVFP4. The timetable and selection criteria for the announced Hub-credit awards are also not provided in the source material.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Community Projects
Developers can use the published integrations and recipes to test Beam for inference, fine-tuning or reinforcement learning. Reflection and Hugging Face say eligible projects can be shared on Beam’s Hugging Face page for consideration for the $10,000 in Hub credits. The materials provided do not state when awards will be decided.
Further reporting and independent evaluations could establish how Beam performs on coding, tool use and long-context tasks, and what hardware is needed for different formats and workloads. Until those results are available, the release supports conclusions about Beam’s stated design, license and integrations, but not its comparative performance.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Beam?
Beam is a text-only, 501-billion-parameter mixture-of-experts model released by Reflection on Hugging Face under the Apache 2.0 license. The company says it is intended for coding and agentic tasks.
How many parameters does Beam use for each token?
Reflection says 23 billion parameters are active per token, with eight experts selected from 256 in each mixture-of-experts layer. This does not mean the full model weights can be omitted from deployment memory.
Can developers run Beam themselves?
The model is available under the Apache 2.0 license and has listed support for Transformers, vLLM, SGLang and Reflection’s Rust-based harness. Its BF16 checkpoint is about 1 terabyte, so self-hosting requires substantial infrastructure; the source does not give a full hardware minimum.
Has Beam’s performance been independently verified?
The supplied release materials do not include independent benchmarks or comparisons with other models. They also do not report measured accuracy, speed or memory use at the stated 1-million-token context length.
What are the Hub credits for?
Reflection and Hugging Face announced $10,000 in credits for selected projects using Beam, including possible demos, quantizations, datasets, reinforcement-learning environments and benchmarks. The supplied material does not specify the award timetable or judging criteria.
Primary source: Hugging Face · via ThorstenMeyerAI.com
