Beam, the US AI model likened to DeepSeek, secures billion-dollar funding and eyes corporate adoption

Beam, a new artificial intelligence system developed in the United States, has been dubbed the “Western DeepSeek” by investors. This nickname references the Chinese company DeepSeek, which gained prominence by offering a competitive model at a lower cost. Beam shares a similar goal: to deliver performance comparable to Chinese rivals while using less computational power.

What is Beam and who is behind it

Beam is the first model from Reflection AI, a New York-based company founded in 2024 by Misha Laskin and Ioannis Antonoglou, former researchers at Google DeepMind. The startup raised $4.6 billion from investors such as Nvidia, Sequoia, and Lightspeed, and is valued at approximately $25 billion.

It is an open-weights model, meaning the trained parameters are available for companies and governments to download, adapt, and run the system on their own infrastructure. According to Reflection, the weights will be published later this month.

How Beam works

The model has 501 billion parameters, which are the internal variables adjusted during training. It uses a mixture of experts architecture: only 23 billion parameters are activated to process each piece of text, reducing operational costs.

Beam works exclusively with text and has a context window of 1 million tokens. A token is the unit into which the model divides text, and the window indicates how much content it can process at once, such as contracts, financial reports, or entire codebases.

The primary focus is on programming, multi-step reasoning, and AI agents, which are systems that execute sequences of tasks with minimal supervision.

Test results

Reflection published the results. The company claims that Beam performs on par with Z.ai’s GLM-5.2 from China, using three to four times less computational power during inference, the stage where the model responds to requests. It also comes close to Alibaba’s Qwen 3.8-Max.

There are caveats. Moonshot’s Kimi K3 still leads in raw capacity, and GLM-5.2 has already been replaced by newer versions. Additionally, Beam does not yet appear in the Artificial Analysis index, which compiles independent tests.

Use in companies and governments

The primary focus is on corporate use. A bank or government agency could run Beam on its own chips, using private data, without sending information to an external vendor. According to international press reports, such a test is currently underway in South Korea.

Within a technology team, the model could be used, for example, to review code or automate steps in an internal process.

What to expect moving forward

With the publication of the weights, researchers and companies will be able to test Beam independently and verify whether the reported results are replicated in real-world tasks, such as code generation and process automation.

The launch also intensifies the competition for open models, which includes DeepSeek, Alibaba, Z.ai, and Moonshot in China, and Meta, Mistral, Nvidia, and Thinking Machines in the West.