What changed
On October 5, Reflection AI, a Brooklyn startup backed by Nvidia and valued at $25 billion, announced Beam, its first open-weight model. Beam is a sparse mixture-of-experts with 501 billion total parameters and 23 billion active per token, built for coding, reasoning, and agent tasks. The company claims Beam matches the performance of Z.ai's GLM-5.2 while using 3 to 4 times less compute. It does not yet beat Moonshot's Kimi K3, but it outscores other US open models like Thinking Machines' Inkling on coding benchmarks. The weights will be released this month under Apache 2.0, which permits commercial use and modification.
What you can do with it this week
Sign up for early access on Reflection's site now. When the weights drop later in October, you can run Beam on your own hardware or a rented GPU cluster. If you have coding tasks, customer-support agents, or internal tools that need reasoning, test Beam against your current API bill. Because it is open-weight, you pay once for compute, not per token. If your team already runs open models like Llama or Mistral, add Beam to your rotation and compare output quality and speed. For tasks where data privacy matters, local deployment means you do not send anything to a third party.
What to keep in your own hands
The model is not public yet, so the claims are Reflection's, not independent. Wait for the technical report and third-party benchmarks before you migrate production work. Open weights mean you own the copy, but running a 501-billion-parameter model is not cheap. Estimate your GPU costs first. If you do not have in-house infrastructure, cloud hosting for a model this size can exceed API pricing unless you have sustained, high-volume use. And remember: open weights are not open training data. You get the model, not the recipe. You can fine-tune it, but you cannot audit how it was built or what it learned. Keep your own records of what data you feed it and what decisions it influences.