Reflection Beam Weights Released: A 501B Coding and Agent Model Under Apache 2.0
What is it?
Reflection AI's Beam is a sparse mixture-of-experts (MoE) model with 501B parameters in total and 23B active per token. It targets coding and agent work, and the weights are Apache 2.0. The 8-bit and 4-bit versions fit on one eight-GPU server of H100, H200 or GB300 chips; the full 16-bit version needs tens of GPUs. The hosted service is free during the public beta.
Why does it matter?
Reflection says Beam matches GLM-5.2 on advanced reasoning tests while using 3 to 4 times less inference computeVendor claim. That's worth a look if you self-host agents and are pricing it out. But its Terminal-Bench 2.1 score of 80.1% trails DeepSeek V4.1 Flash and Kimi K3Vendor claim, and fewer tokens doesn't automatically mean a cheaper task.
Maturity
GAWeights are out, but there's no price yet and no independent evaluation.
How to track it
- Independent evaluation: look for real cost-per-finished-task results from VentureBeat or Artificial Analysis.
- Pricing: see whether the Beam page on OpenRouter gets a price.
- Quantization: watch for quality reports on community conversions such as the GGUF builds on Hugging Face.