Reflection’s Beam is available on Hugging Face under Apache 2.0, with examples for running and adapting the coding model. The release includes supervised training and reinforcement-learning paths alongside inference software support. Its expert-based architecture activates about 23 billion parameters per token out of 501 billion, while all expert weights remain in memory. The model takes text and returns text.
Reflection, an artificial intelligence developer, has released Beam’s weights through the model-sharing platform Hugging Face under Apache 2.0. Beam accepts text and returns text, with coding and tool-using agents among its intended tasks. Hugging Face worked with Reflection on examples for running the model, supervised fine-tuning and reinforcement learning, which adjusts behavior using feedback.[1], [2]
Beam activates about 23 billion of its 501 billion parameters for each token, a unit of text. After its first 3 dense layers, 49 layers use a mixture-of-experts structure. Each expert layer selects 8 out of 256 experts. All expert weights remain in memory. Most attention layers examine a sliding window over the preceding text, while 9 layers can attend to the complete context.[1]
Initial serving support includes Transformers, SGLang and vLLM, software for loading and running models. A separate execution environment written in Rust accompanies Beam. The release also supplies supervised-training examples using TRL, reinforcement-learning examples using Miles and smaller LoRA adapters for customization. An option retains reasoning within the conversation context, alongside a dedicated grammar for tool calls.[1]
