Reflection Beam: a 501B open model under Apache 2.0

Reflection, a new US lab, has released its first model: Beam. According to the announcement on the Hugging Face blog, the model is open and ships under the Apache 2.0 license. According to the post, teams can run the model on their own hardware or cloud and customize it for their own use cases.

What does the announcement say?

• Beam is a mixture-of-experts (MoE) model with 501B parameters, of which about 23B are active per token.
• It takes text as input and has a context window of 1M tokens.
• Its focus is writing code and working as an agent.
• Hugging Face also writes that Reflection will give out $10,000 in Hub credits to the best projects built with it.

What it means for a small business

This is not built for a restaurant's or a boutique's computer. Hugging Face recommends four B200 GPUs for long contexts and concurrent requests. Still, the news matters for two reasons:

• Choice. An open-license model opens another route for teams that want to run it on their own infrastructure.
• A reason to ask. When you buy a service that uses AI, it is reasonable to ask which model runs it.

Our practical advice is simple: do not try to install Beam yourself. But if you pay for a service that uses AI on your website, menu or ads, ask your provider which model runs it and where your data goes.

Source: Hugging Face blog, Beam announcement. This note relies on the statements in that announcement; we did not measure the model's performance.