Reflection releases Beam, an open model with 501 billion parameters

Reflection AI, a startup in the US with Nvidia as a backer, made Beam, its first model. Beam has 501 billion parameters and is for code and agent tasks. A waitlist is open for developers. Reflection tells that the weights come in October with the Apache 2.0 license, which gives the right to use it in products.

How it works

Beam is a mixture-of-experts model. For each token (a small piece of text), it uses only 23 billion parameters. Reflection made it with 23.8 trillion tokens. It then used reinforcement learning (training with a score for each result) on code and agent tasks.

The numbers

On the Artificial Analysis Intelligence Index, Beam has a score of 30. The model of Meta, Muse Spark Open, has 39. Beam used about 26,000 output tokens for each task. It makes 228 tokens each second.

What is not known

Reflection did not give its training data. Its claim of less computation is an estimate, not a measure of cost. The test puts some open models from China above Beam.

Insight: Nvidia can pay for the open model that uses its hardware

Nvidia put $800 million in Reflection. Reports from 10 October tell about talks in which Nvidia can put in more money or get the company. The source tells that none of its Nemotron models is as good as Beam. This reads as a move to back a strong open model, because the source tells that open models give more work to Nvidia hardware.