Reflection AI publishes Beam technical report for its first open-weight model

Reflection AI released a detailed technical report, about 156 pages, for Beam, its first open-weight model built for coding, reasoning, and agentic tasks. Beam is a sparse mixture-of-experts model with 501.3 billion total parameters but only 23.4 billion active at a time, spread across 52 layers and 256 experts. Reflection says this design needs three to four times less inference compute than many models in its class. The model was pretrained on 23.8 trillion tokens in under four weeks using 6,144 NVIDIA GB300 GPUs, and supports a 1 million token context window through a two-stage midtraining process. Post-training reinforcement learning generated more than 100 million rollouts on 10,500 GB300 GPUs over another four weeks, and Reflection says capability is still improving as more RL compute is applied. The model uses FP8 and NVFP4 quantization along with a small 1.93 billion parameter speculative decoding model called DFlash 2 to speed up inference. On safety testing, Beam refused or redirected 88.9 percent of harmful requests while still complying with 81.5 percent of harmless ones, and was rated Limited across all four systemic risk categories under Reflection's own Open Safety Framework.

Why it matters

It is a rare case of a frontier-scale open-weight model shipping with a full technical report and documented safety testing, giving outside researchers real detail to evaluate instead of a marketing page.

Source: Reflection AI