Beam 1.0 501B A23B
Reflection AI · 2026-10-16 · 501.3B parameters
Beam 1.0 is Reflection's first open-weight model: an instruction-tuned, text-only sparse mixture-of-experts reasoning model for coding and long-horizon agentic work, with 256 routed experts (8 active per token) across 52 layers (3 dense, 49 MoE). Reflection announced it on October 5, 2026 as a waitlisted early-access API and released the weights on Hugging Face under Apache 2.0 on October 16, 2026, together with FP8 and NVFP4 numerics of the same model and a self-serve API that serves this identifier. The BF16 checkpoint contains exactly 501,267,612,352 parameters; Reflection reports about 23.4B active parameters per token (rounded). Reasoning is always produced on an analysis channel, and the chat template exposes a reasoning_effort budget with labels minimal, low, medium, high, xhigh and max (default max). Reflection states a 1M-token context and a June 30, 2026 knowledge cutoff; pretraining used 23.8 trillion tokens before mid-training and RL, so cumulative training exposure is not stated.