Reflection, a new US-based lab, released Beam on October 16, its first model and an open one. Beam is a mixture-of-experts (MoE) language model with 501 billion parameters in total, of which 23 billion are active for each token, and it ships under the Apache 2.0 license.
What Beam offers
According to the Hugging Face launch post, Beam takes text in and produces text out, with a context window of 1 million tokens. It is aimed at coding and agentic work, and it comes with its own open-source, Rust-based harness that Reflection says brings out the best performance from the model. The architecture uses 52 layers and 256 experts per MoE layer, with eight experts routed per token.
Checkpoints are published in BF16, FP8 and NVFP4 formats. Beam supports five reasoning settings, from “low” to “max”, and can keep its reasoning in context across turns. Day-zero support is available in transformers, vLLM and SGLang, and Hugging Face says it worked with Reflection on recipes for inference, supervised fine-tuning and reinforcement learning.
Where to run it
AWS says Beam can be deployed on Amazon SageMaker JumpStart as a real-time endpoint, with other deployment options described in its own post. Reflection is also offering $10,000 in Hub credits to the best projects built with Beam.
Why it matters
Frontier-scale open weights under a permissive license are rare. Teams that want to host and customize a large coding model on their own infrastructure now have another option, though independent benchmark results are still to come.