Ex-DeepMind founders' Reflection AI unveils Beam, a 500-billion-parameter open model

Reflection AI, a US start-up founded by two former Google DeepMind researchers, unveiled its first AI model, called Beam, on Monday, October 5, 2026. The company describes Beam as a frontier "open-weight" model, meaning the trained model itself will be published so that businesses and governments can download it and run it on their own computers. Reflection says Beam has about 500 billion parameters, of which about 23 billion are active for each token, and that it is built for coding and "agentic" tasks, where AI carries out multi-step jobs. The company plans to publish the weights under the permissive Apache 2.0 licence, with a technical report and model card, later in October. At the time of the announcement the model was still in final safety testing, known as red-teaming.

Reflection's pitch is aimed at a gap in the market. Many organisations, from banks to government departments, want powerful AI they can run inside their own systems, so sensitive data never leaves their control. In recent times, some of the strongest open-weight models have come from Chinese labs, and some Western buyers are wary of using them. Reflection is positioning Beam as an American alternative.

The founders' story also matters for the start-up world. Both co-founders worked at Google DeepMind before starting the company, and Reflection has drawn backing from some of the biggest names in technology and finance. Its first product is not an app but a foundation model, the kind of large, expensive system usually built only by the biggest labs. Its claim of needing far less computing power than rivals, if it holds up, is central to whether a start-up can compete at that level.

The caution is that, for now, nearly everything rests on the company's own numbers. Independent testing will begin only once the weights are out. Outside reviewers have already noted that some of the compute estimates leave out parts of the cost of running a model in practice.

For Indian founders and developers, an additional strong open-weight model means more choice of foundations to build products on, without paying per use for a closed service. Indian companies building AI tools for sensitive sectors, or wanting to keep data in the country, are among those who could benefit.

Short-term: Developers are waiting for the weights and technical report later in October to test Reflection's claims independently.

Long-term: If Beam performs as claimed, it gives companies and governments another open model to build on, and shows a start-up can compete with large labs at the frontier.

Who is affected: AI developers and enterprises seeking self-hosted AI, governments concerned about data control, rival AI labs, and investors backing foundation-model start-ups.

Questions and Answers