Back to Model
R

Reflection Beam

Model Oct 16, 2026

Reflection's Apache-2.0 text model with 501B total and 23B active parameters, released in BF16, FP8 and NVFP4 formats.

Visit primary link 3 sources attached
Reflection's English Beam production inference dashboard illustration showing prefill, decode, cache and throughput fields powered by SGLang.
First frame of Reflection's published production inference dashboard animation. Counters and throughput are the company's displayed figures, not an AITrending measurement or a promise for a user's hardware. Full frame preserved. Source: Reflection
self-hostedcoding-agentsagent-ecosystem

What happened lately

Reflection Beam timeline

Reflection made Beam’s weights publicly available on October 16, 2026, following its October 5 preview announcement. The model card describes a text-only sparse mixture-of-experts model with 501B total parameters, 23B active per token and a one-million-token context window. BF16, FP8 and NVFP4 checkpoints are released under Apache-2.0.

Hugging Face documents Transformers, vLLM and SGLang support. Its deployment guide estimates roughly 1TB for BF16 weights and around 280GB for NVFP4; active parameters do not describe the total storage requirement. Reflection also presents Mirror CLI as its companion agent harness and lists other supported harnesses. Model benchmark and efficiency results remain attributed to their publishers rather than independently reproduced by AITrending.