Welcome Beam: a 501B open model you can own, customize, and scale

本文为原文前 6,000 字符的节选翻译，完整内容请查看原文。

Title: Welcome Beam: a 501B open model you can own, customize, and scale

Welcome Beam: a 501B open model you can own, customize, and scale. Published October 16, 2026.

欢迎使用 Beam：一个您可以拥有、定制和扩展的 501B 开源模型。发布于 2026 年 10 月 16 日。

Beam by Reflection is out on Hugging Face! It is the first model from the new US lab, Reflection, and it’s open! It’s a mixture-of-experts (MoEs) model with 501B parameters (23B active per token), supports text as input, has a context window of 1M tokens, and is released under the Apache 2.0 license. Beam is focused on coding and agentic use cases.

Reflection 推出的 Beam 现已在 Hugging Face 上线！这是来自美国新实验室 Reflection 的首个模型，并且是开源的！它是一个拥有 501B 参数（每个 token 激活 23B）的混合专家模型（MoE），支持文本输入，拥有 100 万 token 的上下文窗口，并以 Apache 2.0 许可证发布。Beam 专注于编程和智能体应用场景。

It is for teams that want a frontier-scale model on their own hardware, infrastructure, or cloud, and want to customize it to their own use cases. Beam ships with its own dedicated open source Rust based harness that delivers the best performance from the model. Beam has a native transformers implementation, as well as downstream libraries like vLLM and SGLang.

它专为那些希望在自己的硬件、基础设施或云端运行前沿规模模型，并希望针对特定用例进行定制的团队而设计。Beam 自带专用的开源 Rust 驱动程序，可提供最佳的模型性能。Beam 拥有原生的 transformers 实现，以及 vLLM 和 SGLang 等下游库。

TL;DR: 501B-parameter MoEs LLM with 23B active per token. 52 layers, 256 experts per MoE layer, 8 routed per token. Text in, text out with context window of 1M tokens. Checkpoint formats: BF16, FP8, and NVFP4. Reasoning that can stay in context across turns (preserve_thinking), tool calling with a dedicated grammar.

简而言之：501B 参数的 MoE 大语言模型，每个 token 激活 23B。52 层，每个 MoE 层 256 个专家，每个 token 路由 8 个专家。文本输入输出，上下文窗口为 100 万 token。检查点格式：BF16、FP8 和 NVFP4。支持跨轮次保持上下文的推理（preserve_thinking），以及带有专用语法的工具调用。

Reflection is going to give out $10,000 Hub credits to the best projects that use Beam. All you have to do is make a thing, and share it on the Hub. You could make a demo that uses the Beam model on Spaces, quantize it for your favorite hardware, generate a synthetic dataset, build an RL environment, or benchmark Beam and share the results.

Reflection 将向使用 Beam 的最佳项目发放 10,000 美元的 Hub 积分。您只需制作一个项目并将其分享到 Hub 上即可。您可以制作一个在 Spaces 上使用 Beam 模型的演示，为您的硬件进行量化，生成合成数据集，构建强化学习环境，或者对 Beam 进行基准测试并分享结果。

What is Beam? Beam is a text-only model built for long-context (1M context length) reasoning and agentic work. It is a sparse MoE model, where each token is computed by a small fraction of the network, so it runs at the speed of a much smaller model while keeping the capacity of a large one.

什么是 Beam？Beam 是一个纯文本模型，专为长上下文（100 万长度）推理和智能体工作而构建。它是一个稀疏 MoE 模型，每个 token 仅由网络的一小部分计算，因此它在保持大型模型容量的同时，运行速度却像小模型一样快。

The decoder has 52 layers with a hidden size of 5,120. Most layers use sliding-window attention over the previous 512 tokens; every sixth layer, and the last one, attend to the full context. That gives 43 sliding-window layers and 9 global ones.

解码器有 52 层，隐藏层大小为 5,120。大多数层对之前的 512 个 token 使用滑动窗口注意力机制；每隔六层以及最后一层会关注整个上下文。这构成了 43 个滑动窗口层和 9 个全局层。

The first 3 layers use a dense SwiGLU MLP. The other 49 replace it with a Mixture-of-Experts block of 256 experts, each a SwiGLU MLP with an intermediate size of 2,560, or about 39M parameters. The experts hold 98.4% of the model’s parameters. The router follows DeepSeek-V3’s group-limited scheme.

前 3 层使用稠密 SwiGLU MLP。其余 49 层则替换为包含 256 个专家的混合专家块，每个专家都是一个中间大小为 2,560（约 39M 参数）的 SwiGLU MLP。专家占据了模型 98.4% 的参数。路由器遵循 DeepSeek-V3 的组限制方案。

The residual connections are the most distinctive part. A standard transformer adds each sublayer output to the stream with weight 1, so the stream grows with depth. Beam counts the input embedding as step 1. At step k it adds the sublayer output with weight 1/sqrt(k) and keeps the existing stream with weight sqrt((k-1)/k).

残差连接是最独特的部分。标准的 Transformer 以权重 1 将每个子层的输出添加到流中，因此流会随深度增加。Beam 将输入嵌入视为第 1 步。在第 k 步，它以 1/sqrt(k) 的权重添加子层输出，并以 sqrt((k-1)/k) 的权重保留现有流。