Skip to content- Sources: primary
- Summary: Beam is a text-only mixture-of-experts model with 256 experts per layer, 8 routed per token, and hybrid sliding-window and global attention. Reflection ships it under Apache 2.0 in BF16, FP8 and NVFP4 checkpoints, with a DFlash2 speculative decoding drafter and a Rust-based harness. Hugging Face says the NVFP4 checkpoint is about 280 GB and recommends four B200s, and the full 1M-token KV cache takes about 36 GiB per sequence.
- Why it matters: Teams that want a permissively licensed frontier-scale coding model on their own hardware get another option with ready serving and fine-tuning recipes.