
Beam from Reflection is now live on Hugging Face! This is the first model from Reflection, a new US-based lab, and it is completely open source! It is a Mixture of Experts (MoE) model with a total of 501B parameters (23B active per token), supports text input, features a 1M token context window, and is released under the Apache 2.0 license. Beam focuses on code generation and agent use cases. It is tailored for teams looking to run frontier-scale models on their own hardware, infrastructure, or cloud, and customize them for their specific needs. Beam also comes with a dedicated open-source Rust harness designed to extract peak performance from the model.
Beam comes with a native transformers implementation and is supported by downstream libraries like vLLM and SGLang. We collaborated closely with the Reflection team to deliver inference, SFT (Supervised Fine-Tuning), and RL (Reinforcement Learning) workflows so you can build on top of this model to suit your needs. Alongside day-0 support across transformers, SGLang, and vLLM, this post covers how to run it, how to handle post-training, and what we built with it.
TL;DR
- 501B parameter MoE LLM with 23B active parameters per token. Features 52 layers, 256 experts per MoE layer, routing to 8 experts per token.
- Text in, text out, with a 1M token context window.
- Weight formats: BF16, FP8, and NVFP4.
- Supports preserving reasoning in context across turns (
preserve_thinking), along with tool calling using dedicated syntax. The model offers five reasoning effort configurations: “low”, “medium” (default), “high”, “xhigh”, and “max”.- Tokenizer includes built-in response templates.
- Speculative decoding draft model based on DFlash2.
- Day-0 library support: transformers, SGLang, vLLM. Full-parameter RL supported via Miles.
- Ships with a custom Rust-based harness.
- Apache 2.0 license.
Build with Beam
Reflection is offering $10,000 in Hub credits for standout projects built using Beam. All you need to do is build a project and share it on the Hub:
- Build a demo on Spaces powered by the Beam model
- Quantize it for your preferred hardware