Reflection unveils Beam: Open AI model that uses far less computing power

The US AI company Reflection AI has unveiled its first open model, Beam. The company was founded by two former researchers from Google DeepMind and has NVIDIA as an investor.

Open-weight means that the model's "weights" will be published. The weights are the numbers the model learned during training. Others can then download the model themselves, run it on their own servers and adapt it. Reflection says the weights will come under the Apache 2.0 license, which allows broad use, including in companies. That is to happen later in October. Until then, selected users can get early access via a waitlist.

Large model, but only part is in use at a time

Beam has 501 billion parameters in total. Parameters are the settings an AI model adjusts during training. The model is built as a so-called Mixture-of-Experts. That means it consists of many smaller "experts," and only some of them work on each answer. In Beam's case, it is 23 billion parameters at a time. That is why the model can be cheaper to run.

The model can read up to 1 million tokens at a time. That roughly corresponds to several thousand pages of text. It is especially aimed at programming, reasoning and agents, that is, AI that carries out multiple steps of a task on its own.

Training with enormous computing power

According to Reflection, pretraining, the first big training phase, used 6,144 NVIDIA GB300 chips in under four weeks. This was followed by reinforcement learning, where the model practices by trying things out and getting feedback. Here the company used 10,500 GB300 chips for four weeks.

For that phase, more than a billion so-called sandboxes were created. A sandbox is an isolated test environment where the AI model can, for example, run code without doing harm. Up to 170,000 could run at the same time. We have not been able to confirm the daily figure of 46.4 million mentioned in the headline.

What the results say

Reflection reports, among other things, 77.2 percent on the programming test SWE-Bench Pro v2-Hard and 97.8 percent on the math test AIME 2026. These are the company's own figures, and they have not yet been tested independently. The company says Beam performs on par with larger open models such as GLM 5.2 and is approaching Qwen 3.8-Max on coding tasks.

What it means

Many of the strongest open models today come from China. Beam is an attempt to give Western companies an open alternative they can run themselves. If the claim about lower computing power holds, it could make advanced AI cheaper to use. That will be settled when the model is released and others can test it.

What it means for you

If you or your workplace has IT people who want to run AI on their own servers, Beam could become a cheaper and more open choice than the closed services. As an ordinary user, you will not notice it right away, but cheaper AI may eventually lead to better and cheaper tools for, for example, programming. If you want to be early, you can sign up for the waitlist; otherwise you can wait until late October, when others will have tested whether the promises hold.