Reflection releases Beam, a 501B-parameter Apache 2.0 MoE model for coding agents with 1M-token context

Reflection’s first model, Beam, is now on Hugging Face as an open 501B-parameter mixture-of-experts LLM with 23B active parameters per token, text-only I/O, a 1M-token context window, and an Apache 2.0 license. The model targets coding and agentic workflows, with day-0 support for transformers, vLLM, and SGLang, plus BF16, FP8, and NVFP4 checkpoints and recipes for inference, supervised fine-tuning, and reinforcement learning.

For teams weighing self-hosting versus API use, Beam offers customization and open licensing, but the article says all 256 experts must remain in memory, so deployment still needs 501B-scale capacity even though only 23B parameters run per token. FP8 and NVFP4 checkpoints may reduce hardware pressure, but the reported benchmarks come from Reflection and the article does not provide independent latency, cost, or production reliability data.