The full story · 1 min

Reflection AI was founded by two former Google DeepMind researchers, Misha Laskin and Ioannis Antonoglou. It has raised about $4.7 billion, TechCrunch reports, and its last round valued it at $25 billion before the new money. Beam is its first model.

What Beam is

Beam is a mixture-of-experts model with 501 billion parameters, of which 23 billion are active for each token, a word or part of a word. Reflection says it trained it on 23.8 trillion tokens of text.

It reads only text, TechCrunch notes, and can handle up to 1 million tokens at once. Reflection says the model is “undergoing final red-teaming and evaluations”.

The weights, a technical report and a model card will come “later this month”, under the Apache 2.0 licence, Reflection says. That licence lets anyone use and change the model, including for business. Until then, people can only sign up for early access.

Sources12

Map
Close to GLM, behind Kimi and Qwen.

Terminal-Bench 2.1 scores from Reflection’s own announcement. No outside test yet.

Reflection’s claims

On advanced reasoning tests, Beam scores about the same as GLM-5.2 from the Chinese lab Z.ai “while using 3–4× less inference compute”, Reflection says. It also says Beam beats the leading Western open models.

Its own table is more mixed. On Terminal-Bench 2.1, a coding test, Reflection puts Beam at 80.1, against 81.0 for GLM-5.2, 86.6 for Qwen 3.8-Max and 88.3 for Kimi K3.

TechCrunch notes that no one outside the company has checked the efficiency claim. Artificial Analysis, which tests models independently, had no page for Beam on October 5.

Sources12

A US answer to Chinese labs

Semafor calls Beam an open-source Western answer to Chinese labs, whose free models have been widely used by developers. Fortune describes it as a US-based open model built to compete with China.

Reflection has deals for computing power with SpaceX and Nebius worth more than $7 billion together, TechCrunch reports. They give it access to Nvidia GB300 chips through 2029.

Sources234

Sources

Every fact in this story comes from the sources below. Open them to check our work.

  1. 1
    Primary source · October 5, 2026Introducing Beam Reflection AI
  2. 2
  3. 3
  4. 4
Corrections
  • We first wrote that Beam uses 23 billion parameters for each word. Reflection gives 23 billion active parameters for each token, and a token is often only part of a word.

AI desk: Open models · published October 5, 2026, 22:09 CEST. This story was generated by AI systems under the editorial rules of Silicon AI News and checked against the sources it links. No person read it before it was published. The site’s owner and editor: Y. D. Daniel, Budapest, Hungary. Found a mistake? Email editor@siliconainews.com. How we report

We count only the page and your answer.Thank you. We counted only this page and your answer.
More: words to know, tags and how to cite this story
Words to know
Open-weight model
An AI model whose trained numbers anyone can download and run.
Parameter
One of the many numbers inside an AI model that it learns in training.
Mixture of experts
A model design that uses only some of its parts for each token, a word or word piece, to save computing.
Basics explainedWhat is an AI model? What is the difference between open and closed models?

Reflection unveils Beam, an open model to rival China’s labs. Silicon AI News, October 5, 2026. https://siliconainews.com/stories/reflection-beam-open-model