01 What happened
On 16 October 2026, Reflection released its first open model, Beam. Beam is a sparse MoE LLM with 501B parameters (23B active per token), 52 layers, 256 experts per MoE layer (8 routed per token), and a 1M-token context window. It targets coding and agentic workflows.
02 Key details
- 501B parameters (23B active), 52 layers, 256 experts per MoE layer, 8 routed per token. Apache 2.0.
- Benchmarks provided by Reflection: SWE-bench Verified 81.6%, AIME 2026 97.7%, GPQA Diamond 90.6% (BF16).
- Hardware: NVFP4 checkpoint (~280 GB) requires Blackwell GPUs (4x B200 or 2x B300); FP8 for Hopper/Ampere.
- Day-0 support in Transformers, SGLang, vLLM. Post-training recipes: LoRA SFT with TRL and GRPO with Miles.
03 Why it matters
Teams can own, customize, and deploy a frontier-scale open model on their own hardware, with dedicated recipes for inference, SFT, and RL.
04 Who it matters to
AI developers, AI agent developers, machine learning engineers, AI infrastructure engineers and artificial intelligence researchers.
Original source