Reflection AI Releases Beam, a 501B Open-Weight Model

Reflection AI has published the weights for Beam, a sparse Mixture-of-Experts model tailored for programming, reasoning, and autonomous agent workflows. Distributed under an Apache 2.0 license via Hugging Face, the model features 501 billion total parameters but activates only 23 billion per token during inference. Pretrained on 23.8 trillion tokens using 6,144 NVIDIA GB300 NVL72 GPUs over less than four weeks, Beam finished pretraining with 92.3 percent cluster goodput.

To refine its agentic behavior, Reflection executed a massive post-training campaign. The run generated more than 100 million reinforcement-learning rollouts using 10,500 NVIDIA GB300 GPUs over four weeks, maintaining an average of 110,000 concurrent rollouts and deploying over 1.3 billion total sandboxes. Beam introduces an adjustable reasoning-effort setting, allowing developers to allocate larger token budgets for improved accuracy. The company reports that Beam delivers reasoning performance comparable to GLM-5.2 while utilizing three to four times less inference compute.

Alongside the model weights—available in BF16 at roughly 1TB, FP8 at 500GB, and NVFP4 at 250GB—Reflection released Mirror CLI, the Rust-based terminal agent harness used during training. The release includes a speculative-decoding drafter model to accelerate serving, as well as self-distillation recipes using LoRA adapters. The model features day-zero integration with over 40 ecosystem partners, including AWS, Azure, Google Cloud, Databricks, vLLM, SGLang, Unsloth, Axolotl, and NVIDIA Dynamo, with enterprise options provided by Dell and HPE for private deployments.