• Sources: primary, primary, discussion
  • Summary: Reflection released Beam 1.0 on 2026-10-16 as open weights under Apache 2.0, with BF16, FP8 and NVFP4 checkpoints, about 23B active parameters per token and a context window of roughly 1M tokens. Google Cloud published day-0 serving recipes for llm-d with vLLM and NVIDIA Dynamo with SGLang, and reports up to 5x throughput from KV-cache-aware routing, prefill and decode disaggregation and KV offload. Google says those figures come from customer benchmarking.
  • Why it matters: Teams running self-hosted coding agents get a permissively licensed frontier-size model, with public Kubernetes recipes for multi-turn agent serving.