Reflection's Beam: a 501B MoE open model with 1M context, Apache 2.0, agentic focus
mervenoyann · x · 2026-10-16
Beam, the first model from new US lab Reflection, is out on Hugging Face as a truly open release aimed at coding and agentic workloads:
- Specs: 501B-parameter MoE with 23B active per token; 52 layers, 256 experts per MoE layer with 8 routed per token; text-in/text-out with a 1M-token context window; Apache 2.0 license
- Checkpoints: BF16, FP8, and NVFP4
- Capabilities: reasoning that persists across turns (preservethinking), tool calling with a dedicated grammar, five reasoning levels (low/medium/high/xhigh/max, default medium), plus a DFlash2 speculative-decoding drafter
- Ecosystem: day-0 support in transformers, vLLM, and SGLang; full-parameter RL via Miles; HF and Reflection published inference, SFT, and RL recipes, and the model ships with a dedicated open-source Rust harness for best performance
Positioned for teams that want a frontier-scale model they can own, customize, and run on their own hardware or cloud.
More from Infra
- vLLM ships Day-0 RL support for Beam in SkyRL, +62% LoRA decode throughput — vllm_project · 2026-10-16
- DFlash2 speeds Beam decode up to 3.03× across benchmarks, vLLM reports — vllm_project · 2026-10-16
- Fused kernels cut GB200 forward latency up to 20% in Beam serving microbenchmarks — vllm_project · 2026-10-16
- Inside Beam: 501B params, only 23B active per token via 8-of-256 expert routing — vllm_project · 2026-10-16
- CoreWeave founders cashed out $3.3 billion before the stock's meltdown — SumitGup · 2026-10-16
- The Verge spent a week with Nvidia's RTX Spark Arm laptops: here's what we learned — tomwarren · 2026-10-16