🔍 Read the full analysis: Build On Welcome Beam, The Customizable 501B Open Model on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
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 has 23 billion active parameters per token and supports a context window of up to 1 million tokens, but the release materials do not provide independent benchmark results or a full hardware-cost breakdown.
Reflection has released Beam, a text-only model with 501 billion total parameters, on Hugging Face under the Apache 2.0 license, as detailed in the original analysis. The company says Beam is designed for coding and agentic tasks, supports a context window of up to 1 million tokens, and is available with integrations for Transformers, vLLM and SGLang. The announcement presents the model’s architecture and tooling, but does not include independent performance results.
Beam uses a mixture-of-experts architecture, with 23 billion parameters active for each token, according to Reflection’s release materials. The model has 52 layers; each mixture-of-experts layer contains 256 experts, and a router selects eight for each token. This setup limits how many parameters are used in an individual token computation, but does not eliminate the need to keep the full set of experts available in memory.
The scale has direct deployment implications. The release lists the BF16 checkpoint at about 1 terabyte in safetensors format. It also lists FP8 and NVFP4 versions, with NVFP4 intended for NVIDIA Blackwell GPUs and FP8 versions available for Hopper and Ampere chips. The materials do not quantify the memory savings or quality trade-offs among these formats, and they do not give a complete minimum-hardware or operating-cost estimate.
Reflection says Beam accepts and produces text and includes tool calling with a dedicated grammar, adjustable reasoning settings, and a “preserve_thinking” mechanism intended to carry reasoning across turns. The release also describes recipes for inference, supervised fine-tuning and reinforcement learning, including full-parameter reinforcement learning with Miles. These are capabilities and support described by the project; the announcement does not supply independent tests of their results.
Deployment Depends on More Than Active Parameters
Beam’s release gives developers access to a model that can be used, modified and redistributed under the terms of its Apache 2.0 license. That may interest teams seeking to adapt a model for internal coding or agent workflows, or to choose between their own infrastructure and cloud hosting. Support in established inference libraries and the published post-training recipes may also make it easier for technically equipped teams to begin testing the model.
However, the difference between 23 billion active parameters and 501 billion total parameters is central to planning deployment. Sparse computation can reduce the work performed for each token, but the model’s experts still need to be stored. The approximately 1 TB BF16 checkpoint is a concrete sign of the scale involved. Lower-precision checkpoints may offer different deployment options, but the supplied materials do not establish their performance, quality, or cost advantages.
Reflection and Hugging Face also announced $10,000 in Hub credits for selected projects using Beam. Eligible submissions may include demos, quantizations, datasets, reinforcement-learning environments or benchmarks shared on Hugging Face. Such work could offer examples of how developers use the model, though submissions and vendor-provided recipes are not substitutes for controlled comparisons against other systems.
high performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Architecture and Release Support
Beam combines sliding-window and global attention, according to the release. Of its 52 layers, 43 attend to a window of 512 prior tokens, while nine can attend across the full context. Reflection presents this hybrid design alongside its stated support for inputs of up to 1 million tokens. The announcement does not report independent accuracy or speed measurements at that length.
The model is described as Reflection’s first model. It is text-only, and the release says it has day-one support in Transformers, vLLM and SGLang, as well as Reflection’s Rust-based harness. Hugging Face says it worked with Reflection on support for running and adapting Beam, including training recipes. The release also says the eight selected experts account for 98.4% of the model’s parameters; that architectural detail does not change the need to account for the full checkpoint when planning to host it.
The announcement characterizes Beam as open and names Apache 2.0 as its license. Developers should consult the published license and model materials for the precise terms that apply to their intended use. The release’s practical value will depend not only on access to weights and software support, but also on independent evidence about task performance and the resources needed to run the model.
“501B-parameter MoEs LLM with 23B active per token.”
— Reflection and Hugging Face, in the Beam release materials
large memory server for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Performance and Hardware Still Need Testing
The supplied announcement contains no independent benchmark results, comparisons with other models, or evidence establishing how Beam performs on coding and agentic tasks. It also provides no measurements of quality, latency or memory use during long-context inference. Those omissions mean the stated architecture and feature list cannot, on their own, show how well Beam works in a particular deployment.
Hardware needs will vary with checkpoint format, serving configuration and workload, but the release does not provide a complete minimum-hardware specification or cost estimate. It identifies target GPU generations for some formats and gives the BF16 checkpoint size; it does not quantify the effect of FP8 or NVFP4 on performance or output quality. The model’s stated 1-million-token context is a capability claim in the release, not an independently verified result in the material provided.
The credits announcement also leaves practical details open. The supplied information does not give a judging timetable or a date for award decisions. It is not yet clear which community evaluations will emerge, how comparable they will be, or whether they will include the hardware, prompts and task definitions needed for readers to interpret results.
AI inference hardware NVIDIA Blackwell
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Community Evaluations Could Add Evidence
Developers can use Beam’s Hugging Face release, listed library integrations and published recipes to test inference and post-training. Reflection and Hugging Face say qualifying project links will be collected through the model’s Hub page and that selected projects will receive credits in their Hub accounts. The announcement does not specify when submissions close or when winners will be selected.
The next useful evidence will come from evaluations that report task definitions, hardware, checkpoint format and testing methods, particularly for coding, agentic work and long-context use. Comparable results could help teams judge whether Beam’s capabilities justify its deployment requirements. Until those details are available, its performance and operating costs remain open questions.
enterprise AI model deployment hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Beam?
Beam is a text-only model released by Reflection on Hugging Face. The release describes it as a 501-billion-parameter mixture-of-experts model intended for coding and agentic tasks.
How many parameters are active for each token?
Reflection says 23 billion parameters are active per token. Beam has 501 billion parameters in total, so the active count does not represent the memory needed to hold the full model.
What does the release say about Beam’s context window?
The release states that Beam supports a context window of up to 1 million tokens. The supplied materials do not include independent measurements of accuracy, speed or memory use at that length.
What hardware is required to run Beam?
The announcement does not provide a complete minimum-hardware specification. It lists a BF16 checkpoint of about 1 terabyte and identifies NVIDIA Blackwell GPUs for NVFP4 and Hopper and Ampere chips for FP8 versions.
Are independent performance results available in the announcement?
No. The supplied release materials include architecture details and software support, but do not provide independent benchmarks or direct comparisons with other models.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
