Reflection’s Beam is the open front door to a business selling private ‘AI factories’

Reflection AI, a two-year-old Brooklyn startup founded by two former Google DeepMind researchers, unveiled Beam, its first frontier open-weight model, TechCrunch reports. It says Beam matches leading Chinese open models on advanced reasoning benchmarks at a fraction of the compute; those claims have not been independently verified. Beam has 501 billion total parameters with 23 billion active, was pretrained on 23.8 trillion tokens and has a 1 million-token context window (architecture details). For comparison, Z.ai’s GLM-5.2 has roughly 744 billion total parameters with 40 billion active. Reflection says weights and full technical details will be released this month through hyperscalers, neoclouds and open-source libraries.

Chinese labs have set the pace in open weights. A credible Western alternative would widen choices for companies and governments that want to run AI on their own terms, including hospitals, banks and agencies that need to inspect, adapt and host models locally.

Reflection has raised roughly $4.7 billion, per PitchBook, from backers including Nvidia, Sequoia and Lightspeed, at a $25 billion pre-money valuation in its last round. In the summer it signed deals worth more than $7 billion with SpaceX and Nebius for access to Nvidia GB300 chips through 2029. This is a heavily financed company treating open weights as a sales channel.

The business is AI factories: customized local systems institutions build by training Reflection’s models on their own proprietary data. Nvidia’s CEO has championed the idea for years, and the vision also sells many Nvidia GPUs. Axios reported interest from hedge funds and trading firms, and Reflection is testing a sovereign AI factory partnership with Shinsegae Group in South Korea. Selling deployment services is a transparent way to fund frontier training. The risk is a different concentration: if the Western open ecosystem consolidates around a few richly funded players tied to one chip supplier, openness is real at the weights level and narrower elsewhere.

Reflection faces closed labs like Anthropic and OpenAI, Chinese open models, and Western players including Mistral, Meta and Cohere. Its closest US rival may be Inkling from Thinking Machines Lab. Reflection’s own benchmarks show Beam ahead on four coding tests, but Inkling is multimodal and Beam is text-only, so the comparison depends on which capability buyers need. Mistral is making a parallel sovereignty pitch (Le Chonk and ownership).

Two things determine the rest of the pitch: whether the weights ship on schedule and whether outside testers confirm the efficiency claim. More in our startups coverage, Nvidia coverage, funding coverage and research coverage.