🔍 Read the full analysis: Meet Welcome Beam: An Open 501B Model Built For Customization 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 release lists coding and agentic work as target uses and claims a context window of up to one million tokens, but provides no independent benchmark results; its BF16 checkpoint is about 1 terabyte.
Reflection has released Beam, a text-only, 501-billion-parameter mixture-of-experts model, on Hugging Face under the Apache 2.0 license. The company says Beam is designed for coding and agentic tasks and supports a context window of up to one million tokens; the original analysis also notes that the announcement does not include independent performance results.
Beam has 23 billion parameters active per token, according to Reflection’s release materials. The model has 52 layers, with 256 experts in each mixture-of-experts layer; a router selects eight experts for each token. This sparse design limits the number of parameters used in an individual token’s computation, but does not remove the need to host the full set of model weights.
The scale is reflected in the listed checkpoint size: the BF16 version is about 1 terabyte in safetensors format. Reflection lists BF16, FP8 and NVFP4 checkpoint formats, with NVFP4 intended for NVIDIA Blackwell GPUs and FP8 versions available for Hopper and Ampere chips. The release does not quantify the quality or speed trade-offs among those formats or provide a full minimum-hardware or cost estimate.
Beam is text-only and includes tool calling, adjustable reasoning settings and a “preserve_thinking” mechanism described as carrying reasoning across turns. It has day-one support in Transformers, vLLM and SGLang, as well as Reflection’s Rust-based harness. The release also lists recipes for inference, supervised fine-tuning and reinforcement learning, including full-parameter reinforcement learning with Miles.
What Beam Offers Model Builders
The release gives developers access to a large model whose Apache 2.0 license permits use, modification and redistribution subject to the license terms. That may be useful to organizations seeking to adapt a model to internal workflows or choose whether to run it on their own infrastructure or through a cloud provider. Published integrations and post-training recipes offer starting points for teams that have the required compute and engineering capacity.
Beam’s architecture also highlights a practical distinction for prospective users: active parameters are not the same as hosting requirements. The 23 billion parameters used per token describe part of the computation, while all experts must remain available in memory, according to the release. The approximately one-terabyte BF16 checkpoint is a concrete indication of the scale involved, even though other formats may offer different deployment options.
Reflection and Hugging Face also announced $10,000 in Hub credits for selected projects using Beam. Potential submissions include demos, quantizations, datasets, reinforcement-learning environments and benchmarks shared on Hugging Face. Such projects could provide practical examples, but submissions and vendor-provided recipes would not by themselves establish comparative model performance.
high performance GPU server for AI models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Beam’s Architecture and Access
Beam combines sliding-window and global attention, according to the release. It says 43 layers attend to a window of 512 prior tokens, while nine layers can attend across the full context. The one-million-token context length is a stated capability; the supplied materials do not include independent tests of accuracy, speed or memory use at that length.
Reflection describes Beam as its first model, and Hugging Face says it worked with the company on support for running and adapting it. The model is published on Hugging Face with formats and recipes intended for developers to test. The stated license is Apache 2.0; users should consult the license and model documentation for the terms that apply to their use and redistribution.
“501B-parameter MoEs LLM with 23B active per token.”
— Reflection and Hugging Face, in the Beam release materials
Benchmark and Hosting Questions
The supplied announcement does not provide independent benchmarks, comparisons with other models or results on coding and agentic tasks. It also offers no measured quality, latency or memory figures for prompts approaching the stated one-million-token limit. Those gaps make it difficult to judge how the model performs in practice from the release alone.
Deployment requirements will vary with checkpoint format, serving configuration and workload. While the materials give an approximate BF16 checkpoint size and identify intended GPU generations for some lower-precision formats, they do not provide a complete hardware or operating-cost breakdown. The effect of FP8 or NVFP4 on model quality is also not specified. The release does not state a judging timetable or award date for the Hub credits program.
Independent Tests and Hub Projects
Developers can use the published checkpoints, library integrations and recipes to test inference or post-training, subject to their available hardware and the model’s license. 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. No selection date is given in the supplied materials.
The next useful evidence for prospective users would be reproducible evaluations of coding, tool use and long-context tasks, with details about prompts, hardware, checkpoint format and comparison models. Until those results are available, Beam’s stated capabilities and architecture describe what the release offers, not independent proof of its performance.
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 release says 23 billion parameters are active for each token.
Does Beam have a one-million-token context window?
Reflection’s release materials state a context window of up to one million tokens. The supplied announcement does not include independent testing of accuracy, speed or memory requirements at that length.
How much memory does Beam require?
The release lists the BF16 checkpoint at about 1 terabyte. Actual deployment requirements depend on the format, serving setup and workload; the supplied materials do not give a full minimum-hardware estimate.
Where can developers run or adapt Beam?
The release lists support for Transformers, vLLM and SGLang, alongside Reflection’s Rust-based harness. It also describes recipes for inference, supervised fine-tuning and reinforcement learning.
Are there independent results showing how Beam performs?
Not in the supplied announcement. It does not provide independent benchmark results or comparisons with other models, so Beam’s real-world performance on coding, agentic or long-context tasks remains to be established.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
