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

Reflection released the weights for Beam, a sparse Mixture-of-Experts model with 501B total parameters and 23B active per token.

AI Desk · Published October 16, 2026
First look: this report is based on official posts from Reflection, Hugging Face and AWS and has not yet been tested hands-on by Maitz.

Reflection first introduced Beam on October 5, 2026, as an early preview with a waitlist. Today it is shipping what it promised: open weights, a model card, a safety technical report, and more than 40 partners for inference, fine-tuning and agent tooling.

What did Reflection announce with Beam?

According to the company, Beam is built for coding, reasoning and long-horizon agent work. It is text-only, with a 1M-token context window. The sparse MoE design keeps 256 experts per layer and routes 8 of them per token.

Checkpoints ship in BF16, FP8 and NVFP4, plus a DFlash2 speculative-decoding drafter. Day-zero support covers Transformers, vLLM and SGLang, with recipes for supervised fine-tuning and reinforcement learning.

How do you run it?

According to Hugging Face, the BF16 checkpoint is about 1 TB in safetensors form. NVFP4 shrinks the footprint to roughly 280 GB and targets Blackwell GPUs. FP8 suits Hopper-class hardware. The post suggests planning for at least four B200 GPUs for comfortable long-context serving.

Without a GPU farm, you can use hosted APIs. Reflection says its platform API is free to start. AWS published a deployment guide for SageMaker JumpStart and HyperPod. Reflection also lists partners including Azure, Google Cloud, Fireworks, Together, Baseten, OpenRouter and Modal.

Agent tooling includes Reflection’s open-source Mirror CLI, plus tested paths with Pi, OpenCode and Hermes. Users can set reasoning effort to low, medium, high, xhigh or max.

Numbers the company publishes

According to Reflection, its RL campaign produced more than 100 million rollouts on about 10,500 NVIDIA GB300 GPUs over four weeks. Company-reported BF16 scores include roughly 81% on SWE-Bench Verified and about 80% on Terminal Bench v2.1. These are company figures, not independent Maitz tests.

Reflection claims strong inference efficiency versus larger open models, saying Beam is competitive with GLM-5.2 while using less inference compute. It also says larger Chinese models such as Kimi K3 still lead on raw capability.