🔍 Read the full analysis: A 501B Open Model To Shape And Scale: Welcome Beam 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 says it targets coding and agentic tasks and supports up to a 1-million-token context, but its release materials do not provide independent performance results or a full deployment-cost breakdown.
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 1 million tokens; independent performance results were not included in the supplied announcement.
Beam has 501 billion total parameters, with 23 billion active for each token, according to Reflection and Hugging Face’s release materials. In its mixture-of-experts design, each MoE layer contains 256 experts, and a router selects eight for each token. The model has 52 layers in total. This arrangement limits the experts involved in an individual computation, but it does not remove the need to store the full set of model weights when serving the model.
The announcement lists BF16, FP8 and NVFP4 checkpoint formats. Reflection says the BF16 checkpoint is about 1 terabyte in safetensors format. The release identifies NVFP4 as intended for NVIDIA Blackwell GPUs and FP8 versions for Hopper and Ampere chips, but does not give a complete minimum-hardware or cost estimate. Beam has day-one support in Transformers, vLLM and SGLang, alongside Reflection’s Rust-based harness.
The materials describe Beam as text-in, text-out, with tool calling through a dedicated grammar and reasoning settings from low to max. They also list recipes for inference, supervised fine-tuning and reinforcement learning, including full-parameter reinforcement learning with Miles. Reflection and Hugging Face are offering $10,000 in Hub credits for selected projects using Beam, such as demos, quantizations, datasets, reinforcement-learning environments or benchmarks shared on Hugging Face.
The Cost of a 501B Model
Beam’s release gives developers access to a frontier-scale model under a permissive license, allowing use, modification and redistribution subject to the Apache 2.0 terms. That can give organizations more flexibility to adapt a model for internal workflows and choose whether to run it on their own infrastructure or through a cloud provider. The included library integrations and training recipes may also make it easier for teams with sufficient computing resources to begin testing and adapting it.
Its scale still sets a high bar for deployment. The 23 billion active parameters describe the subset used for each token, not the total weights that must be hosted. The roughly 1-terabyte BF16 checkpoint shows why the distinction matters: sparse computation can limit per-token work, while the model’s full expert pool remains a substantial memory requirement. Lower-precision formats offer alternatives, but the announcement does not quantify their quality, speed or memory trade-offs.
The announced credit program may support community experiments and demonstrations. Those projects could provide practical examples of how Beam behaves in particular settings. However, model recipes and selected submissions are not substitutes for comparable, independent evaluations that report the tasks, prompts, hardware and measurement methods used.
high performance AI inference server
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Beam Uses Its Experts
Beam’s MoE structure uses a router to select eight of 256 experts in each MoE layer for a given token. The release says those experts account for 98.4% of the model’s parameters. The design aims to draw on a large pool of learned capacity while using a smaller subset in each token’s computation; the stated active count should not be read as the memory needed to hold the model.
For long inputs, the release describes a hybrid attention design: 43 layers use a 512-token sliding window, while nine layers can attend across the full context. Reflection lists a maximum context of 1 million tokens. That is a capability claim in the release materials; the supplied information does not include independent tests of accuracy, speed or memory use at that length.
The release presents Beam as Reflection’s first model and names Apache 2.0 as its license. Hugging Face says it worked with Reflection on support for running and adapting the model, including the listed inference and post-training recipes.
“501B-parameter MoEs LLM with 23B active per token.”
— Reflection and Hugging Face, in the Beam release materials
As an affiliate, we earn on qualifying purchases.
Performance and Hardware Still Unclear
The supplied announcement does not report independent benchmark results or comparisons with other models, including on coding and agentic tasks. It also provides no measured quality, latency or memory results for using the claimed 1-million-token context. As a result, the release establishes Beam’s stated architecture, features and integrations, but not how it performs against alternatives in comparable tests.
Deployment needs will vary with checkpoint format, serving setup and workload. The BF16 file size is given, and the release names GPU generations for some lower-precision formats, but it does not specify a full minimum-hardware configuration, operating cost or the performance effects of reduced precision. The materials also do not provide a judging timetable for the Hub-credit projects.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Project Awards
Developers can use the Hugging Face release, listed inference libraries and published recipes to test Beam or adapt it through supervised fine-tuning and reinforcement learning. Reflection and Hugging Face say qualifying projects shared on Beam’s Hub page may be selected to receive credits in their accounts. The supplied materials do not state when winners will be chosen.
The next useful evidence for prospective users would include independently run evaluations of coding and agentic tasks, tests at long context lengths, and deployment measurements that specify hardware and checkpoint format. Until those details are available, Beam’s stated scale and feature set can guide experimentation, but users have limited public evidence in this announcement for judging its comparative performance or the resources needed for a particular workload.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Beam?
Beam is a text-only mixture-of-experts model released by Reflection on Hugging Face. The release lists 501 billion total parameters and 23 billion active per token.
What license does Beam use?
The release lists the Apache 2.0 license, which permits use, modification and redistribution subject to its terms.
Does Beam need memory for all 501 billion parameters?
The release says the BF16 checkpoint is about 1 terabyte. Although 23 billion parameters are active per token, the materials say the experts must remain in memory, so the active count is not a measure of the full hosting requirement.
Has Beam’s performance been independently verified?
No independent benchmark results are included in the supplied announcement. It does not report comparative coding results or measured performance at the claimed 1-million-token context length.
What support and project incentives are available?
The release lists support for Transformers, vLLM and SGLang, as well as Reflection’s Rust-based harness and recipes for inference and model training. Reflection and Hugging Face also announced $10,000 in Hub credits for selected qualifying projects, with no award date specified in the supplied materials.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
