Reflection AI announces Beam, its first open-weight AI model

In short: Reflection AI says its new Beam model can deliver strong reasoning results while using less computing power, and it plans to publish the model’s weights this month.

What happened

Reflection AI, a two-year-old startup based in Brooklyn, announced Beam, its first “open-weight” AI model. Open-weight means the company plans to share the model’s weights, which are the saved settings that make the AI work, like the recipe and measurements for a dish.

Reflection says Beam is designed for reasoning, coding, and “agentic” tasks, meaning it can take steps toward a goal instead of only answering one question at a time. Beam is text-only, so it works with written input and output, not images or audio.

The company claims Beam performs about as well as Z.ai’s GLM-5.2, a well known open model from China, on advanced reasoning tests. Reflection also says Beam uses 3 to 4 times less “inference compute,” which is the processing power needed each time someone runs the model (like the electricity cost every time you turn on a machine). These performance claims have not been independently verified.

Reflection says it will release Beam’s weights and full technical details later this month. It also says Beam will be distributed through large cloud providers and newer cloud firms, and it will work with popular open source software tools at launch.

Why it matters

If Beam really can deliver similar results with less computing power, it could lower the cost of using high-end AI. That matters to businesses and governments that want to run AI on their own data, especially those looking for alternatives to closed models from companies like OpenAI and Anthropic.

Source: TechCrunch AI