Reflection AI open-sources RL post-training recipes for Beam 1.0, a 501B MoE with 23B active params

ying11231 · x · 2026-10-21

• Miles and Reflection AI released complete RL post-training recipes for Beam 1.0 (501B-param MoE, 23B active, 1M context), with code, training configs and examples fully open-sourced (radixark/miles#4149, beam-support branch).
• Covers full-parameter BF16 GRPO (96 or 64 H200s) and LoRA training (32 H200s), with SGLang-powered rollouts; the model loads straight from HF via Megatron-Bridge, no offline torchdist conversion.
• The cookbook includes data generation (software-engineering trajectories that pass tests, tool-calling reasoning samples), fine-tuning (LoRA/QLoRA/full), and distillation recipes.
• Architecture highlights: 43 sliding-window attention layers (window 512) plus 9 full-attention layers, SSMax positional scaling, Gated GQA (48 query heads / 8 KV heads), sandwich norms, 256 experts with grouped top-8 sigmoid routing, FP32 router and logit softcap of 30. Weights ship in BF16, FP8 and NVFP4.

Original post →

More from Infra

• FireTrident lets 80 coding agents share one repo, clearing Hono's 958 violations in 62s — dinasaur_404 · 2026-10-21
• xAI said to lean on ACC/AEC and copper as Chinese pluggables stay off-limits — jwt0625 · 2026-10-21
• Packed a16z SF infra event features fal, ElevenLabs, BFL and Krea engineers — venturetwins · 2026-10-21
• Baseten ships DeepSeek v4.1 Flash, Reflection Beam 1.0, Nemotron Super 3.5 VL day-zero with claimed top benchmarks — baseten · 2026-10-21
• Modal Engineers Detail Sticky Serverless Routing for Real-Time AI Inference — blelbach · 2026-10-21
• EdgeKernelBench: LLM-Generated Kernels Hit 64.5% Correct but Only 25.5% Fast on Edge Backends — encodelab · 2026-10-21

Infra channel →