Reflection AI open-sources Beam 1.0 under Apache 2.0; SGLang ships day-0 support
BanghuaZ · x · 2026-10-16
Reflection AI released Beam 1.0, a model built for software engineering and long-horizon agents, with open weights under Apache 2.0 and 40+ ecosystem partners. The release includes weights, tokenizer, and reference implementation; BF16, FP8, and NVFP4 checkpoints plus a DFlash2 drafter; and Mirror CLI, the open-source harness used to train Beam. Weights are on Hugging Face.
SGLang ships day-0 support with deployment recipes for B300, GB300, and H200: low-latency uses TP with the DFlash2 drafter at block size 8, high-throughput disables it, and Blackwell NVFP4 enables EP with FlashInfer MegaMoE. Bundled GSM8K scores cover all 1,319 test problems. Day-0 TPU support also arrives via SGLang-TorchTPU, combining RadixAttention and HiCache for cost-efficient agentic serving.
Related event: Reflection AI Open-Sources Beam 1.0, a 501B-Parameter MoE Model(24 posts)→
More from Infra
- Zephon brings elastically deterministic data loading to TorchTitan with identical batches across GPU counts — josh_wills · 2026-10-17
- vLLM serves Reflection Beam 501B MoE from day 0, hitting 708 tok/s with DFlash2 — woosuk_k · 2026-10-17
- Zephon x TorchTitan: bit-identical global batches when resuming training on different GPU counts — PyTorch · 2026-10-17
- Jensen Huang: AI shifting from retrieval to problem-solving may need a billion times more compute — rohanpaul_ai · 2026-10-17
- Haight Street Compute Market: a one-day trading floor for compute deals in SF — Scobleizer · 2026-10-17
- eigenrobot welcomes a possible local data center, citing rising property taxes — eigenrobot · 2026-10-17