01 / The short version
What happened
Reflection released Beam, a 501B-parameter open-source MoE model on Hugging Face under Apache 2.0. It features 23B active parameters per token, a 1M-token context window, and is optimized for coding and agentic tasks. The model uses a unique residual connection scaling mechanism and hybrid attention to manage memory and long-context performance. It is available in BF16, NVFP4, and FP8 formats, with day-zero support in transformers, vLLM, and SGLang. Reflection is offering $10,000 in Hub credits for top projects using Beam.
See the exact references02 / Key takeaways
What you need to know
- 01
Beam is a 501B-parameter open-source MoE model with 23B active parameters, designed for coding and agentic use cases.
- 02
The model uses a novel residual connection scaling method to maintain constant stream scale and hybrid attention to reduce KV cache growth.
- 03
Beam is available in BF16, NVFP4, and FP8 formats, with inference support in transformers, vLLM, and SGLang, and a $10,000 credit incentive for community projects.
Keep in perspective
What to watch for
Benchmark results are provided by the publisher without independent verification; the performance claims, particularly around NVFP4 quality retention, should be validated against independent testing.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · Hugging FaceRead the original announcementhttps://huggingface.co/blog/reflection-beamAI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works