Reflection AI releases Beam weights: 501B open model takes on Chinese rivals

US start-up Reflection AI has released the full weights of Beam, its first open-weight artificial intelligence model, on the Hugging Face platform under the Apache 2.0 licence, the company said on Friday, October 16, 2026. Beam is a sparse mixture-of-experts model with 501 billion total parameters, of which 23 billion are active for any given word it processes, and it is built for coding, reasoning and long, multi-step agent tasks. Reflection first unveiled the model on October 5 and pitched it as an American alternative to the strong open models coming out of China.

Chinese labs such as DeepSeek have released many of the most widely used freely downloadable models, and developers who want to run a model on their own servers, rather than pay for each request to a cloud service, often turn to them. Reflection has pitched Beam as a US answer. It says Beam matches Z.ai's GLM-5.2 on advanced reasoning while using about three to four times less computing power to produce answers. That claim is the company's own and has not been checked independently; outside evaluations are still under way.

The release comes in a busy week for open models. Meta released its 816-billion-parameter Muse Spark Open model on October 15, a day before Beam's weights went live, and Nvidia also published a new open model in its Nemotron family. The competition matters because open weights let anyone inspect, adapt and run a model privately, which is important for sensitive data in banking, health and government.

The "active parameters" figure is the key to Beam's cost claim. In a mixture-of-experts model, only a slice of the network is used for each piece of text, so a 501-billion-parameter model can run with roughly the computing cost of a much smaller one. Fewer active parameters mean lower bills for anyone hosting it.

Benchmark figures published by the company and by different testers do not always line up, because they use different test suites and versions. Readers should treat headline scores with care until independent groups finish their tests.

For India, the release adds another option for start-ups and government projects that want to build on open models while keeping data inside the country. A model that runs on fewer active parameters can be cheaper to host on the GPU capacity that Indian cloud providers and public compute programmes now offer.

Short-term: Developers can download Beam today, and Reflection says it is supported from day one in common tools such as Hugging Face Transformers, vLLM and SGLang. Expect a wave of independent tests over the coming weeks that will show whether its efficiency claims hold.

Long-term: If Beam performs as claimed, it strengthens the case that US labs can compete with Chinese open models, and could push other American companies to release more weights. It also gives businesses more bargaining power against paid AI services.

Who is affected: AI developers and researchers; companies that want to run AI on their own servers for privacy or cost reasons; cloud and GPU providers that host open models; and rival labs in the US and China competing for the open-model market.

Questions and Answers