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 the model is aimed at coding and agentic tasks, supports a context window of up to 1 million tokens, and can be run or adapted using an open source Rust-based harness and several inference libraries.

Beam is a mixture-of-experts (MoE) model with 23 billion parameters active for each token, according to the release materials. Its 52 layers include 256 experts in each MoE layer, with eight selected per token. This reduces the amount of model computation used for an individual token, but the release notes that all experts must remain in memory. The BF16 checkpoint is about 1 terabyte in safetensors format, so the active parameter count does not describe the memory needed to host the full model.

The release lists BF16, FP8 and NVFP4 checkpoint formats. It says NVFP4 is intended for NVIDIA Blackwell GPUs, while FP8 versions are available for Hopper and Ampere chips. Beam has day-one support in Transformers, vLLM and SGLang, alongside Reflection’s Rust-based harness. The Hugging Face post also describes recipes for inference, supervised fine-tuning and reinforcement learning, including full-parameter reinforcement learning with Miles.

Beam accepts and produces text. Its listed features include tool calling with a dedicated grammar and reasoning settings ranging from low to max. The materials describe a mechanism called “preserve_thinking” for carrying reasoning across turns. 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 or benchmarks shared on Hugging Face.

What Beam Offers Model Builders

Beam gives developers access to a frontier-scale model under a license that permits use, modification and redistribution subject to its terms. That may appeal to organizations that want to adapt a model for internal workflows or choose where to run it, including their own infrastructure or a cloud provider. The inference and post-training recipes could lower the practical barrier to experimentation for teams with the necessary compute and engineering capacity.

The difference between 23 billion active parameters and 501 billion total parameters matters when evaluating deployment. Sparse computation can reduce per-token work, but it does not remove the need to store the experts. The roughly 1 TB BF16 checkpoint illustrates the hardware demands; lower-precision formats offer other deployment options, but the source does not quantify their performance or quality trade-offs.

The announced credits contest may prompt community projects that test Beam in specific settings. Such demonstrations could help others understand its behavior, but project submissions or vendor-provided recipes alone would not establish how Beam performs against other models. That requires comparable evaluations and details about the hardware, prompts and tasks used.

How the Sparse Model Is Built

Beam combines sliding-window and global attention. The release says 43 layers attend to a window of 512 prior tokens, while nine layers can attend to the full context. This hybrid design is part of the model’s approach to long inputs. The stated 1-million-token context length is a model capability claim; the materials provided here do not include independent testing of accuracy or speed at that length.

In 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. This arrangement is intended to preserve a large pool of learned capacity while limiting the number of experts used in each computation. It does not make the full set of weights unnecessary during deployment.

Hugging Face says it worked with Reflection on support for running and adapting the model, including recipes for supervised fine-tuning and reinforcement learning. The source presents Beam as Reflection’s first model and describes it as open. The specific license named for the release is Apache 2.0.

““501B-parameter MoEs LLM with 23B active per token.””

— Reflection and Hugging Face, in the Beam release materials

Performance and Deployment Questions

The supplied announcement does not include independent benchmark results, comparisons with other models, or evidence showing how Beam performs on coding and agentic tasks. It also does not report quality, latency or memory measurements for long-context use. Those details would help users assess the model beyond its stated architecture and feature list.

Hardware requirements will depend on checkpoint format, serving setup and workload. The release gives a BF16 checkpoint size of about 1 TB and identifies target GPU generations for some quantized formats, but it does not provide a full cost or minimum-hardware breakdown. The effect of lower precision on task performance is not specified in the supplied material.

Community Tests and Model Use

Developers can download Beam and use the published recipes and library integrations to test inference or post-training. Reflection and Hugging Face say they will collect links to qualifying projects shared on the model’s Hub page, select projects and distribute credits to winners’ Hub accounts. The source does not give a judging timetable or a date for the award decision.

Further independent evaluations could clarify Beam’s coding performance, agent behavior, long-context reliability and hardware demands. Until those results are available, teams considering deployment will need to test the model against their own tasks and infrastructure.

Where I land

I see Beam’s release as a meaningful access and customization opportunity for teams that can support a very large model. Apache 2.0 licensing, multiple serving integrations and post-training recipes make it easier to investigate than a model available only through a hosted service. The claimed 1-million-token context and agentic focus are also relevant to developers building systems that work across long documents or tool-driven tasks.

The strongest counterargument is that openness and an extensive feature list do not establish practical advantage. The reported 1 TB BF16 checkpoint points to a substantial infrastructure burden, and the source material provides no independent benchmarks to show that Beam’s coding or agent performance justifies that burden. I would revise my assessment upward if reproducible comparisons showed strong task results, reliable long-context behavior and workable costs across the listed deployment formats. Evidence of weak performance or steep operating costs would make the release less compelling for most teams.

Key Questions

What is Beam?

Beam is Reflection’s text-only mixture-of-experts model, released on Hugging Face. The release lists 501 billion total parameters, with 23 billion active per token, and a context window of up to 1 million tokens.

What license does Beam use?

The model is released under the Apache 2.0 license, according to the release materials.

Can Beam run on a smaller model’s amount of memory?

Not based on the active parameter count alone. Beam uses 23 billion parameters per token, but all experts must remain in memory; its BF16 checkpoint is reported to be about 1 TB.

Which software supports Beam?

The release lists day-one support in Transformers, vLLM and SGLang, as well as a dedicated Rust-based harness from Reflection.

How can developers enter the project award?

Reflection and Hugging Face say developers can share a Beam-related Space, model or dataset repository on the model’s Hub page. Selected projects will receive $10,000 in Hub credits; the source does not specify a decision date.

Source: Hugging Face