The mixture-of-experts design, where only a small fraction of the network runs for each token, is the key trade-off. All 256 experts stay in memory, so deploying Beam requires the memory of a 501B model, but it computes each token at roughly the speed of a 23B one. The BF16 checkpoint is about 1 TB in safetensors format, with FP8 checkpoints for Hopper and Ampere chips and an NVFP4 checkpoint for NVIDIA Blackwell GPUs.

Hugging Face, which worked with Reflection on the release, is offering day-zero support in transformers, vLLM and SGLang, plus recipes for supervised fine-tuning and reinforcement learning. Beam also ships with its own open source Rust-based harness that Reflection says delivers the model's best performance. To encourage experimentation, the best projects built with Beam and shared on the Hub will receive $10,000 in Hub credits.

The target audience is teams that want a frontier-scale model running on their own hardware, infrastructure or cloud, and that want to customise it for their own use cases rather than rely on a hosted API.