adtestbench Higgsfield

AI news · Reflection AI · models

Reflection releases Beam’s weights under an open licence

The 501-billion-parameter model for long-context reasoning and agents is now free to download, adapt and use commercially.

By Alexandre S. , 10:30 UTC

Anyone can now download Reflection AI’s first model. On Friday 16 October 2026 the US lab put Beam’s weights, the learned numbers that make up the model, on Hugging Face, Hugging Face’s launch post shows. They come under the Apache 2.0 licence, which allows changes and commercial use. Reflection had presented Beam the week before, SiliconANGLE notes.

Retro-futurist illustration: a striped sunset over a grid horizon under a starry sky, with a rocket motif standing on the horizon.
Drawn by adtestbench from “Welcome Beam: a 501B open model you can own, customize, and scale”,

Beam is big but frugal. It holds 501 billion of those learned numbers, called parameters, with 256 smaller expert networks in most of its layers. Only about 23 billion parameters work on each token (a word or part of one). So it replies about as fast as a far smaller model, Hugging Face explains. It reads up to 1 million tokens at once and targets reasoning and agentic work.

Running it still takes serious hardware. Hugging Face suggests four Nvidia B200 chips for comfortable use.

The release includes recipes for further training and a small drafter model that speeds up replies, SiliconANGLE reports. Builders can win too: Reflection puts up $10,000 in Hugging Face credits for standout Beam projects, the launch post says.

Companies that want a large model on their own servers, with their data kept in-house, have one more option to price.

Earlier on this story: Reflection unveils Beam, a 501B open-weight model