Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost

Reflection AI, a two-year-old Brooklyn startup founded by two former Google DeepMind researchers, unveiled Beam, its first frontier open-weight model. It says Beam matches leading Chinese open models on advanced reasoning benchmarks at a fraction of the compute. Those claims haven’t been independently verified.

The stakes go past one model. Chinese labs have set the pace in open weights. A credible Western alternative changes the choices available to companies and governments that want to run AI on their own terms.

We think a strong Western open-weight ecosystem is a good thing, and not only for geopolitical reasons. Open weights let institutions inspect, adapt and run models locally, which is what hospitals, banks and public agencies often need. If Beam delivers, it widens that option.

But look at who is behind it. 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. This summer it signed deals worth more than $7 billion with SpaceX and Nebius for access to Nvidia GB300 chips through 2029. This is not a scrappy open-source project. It is a heavily financed company betting that open weights are a sales channel.

That bet has a name: AI factories, customized local systems that institutions build by training Reflection’s models on their own proprietary data. Nvidia’s CEO has championed the idea for years, and it’s worth saying plainly that the vision also sells a lot of Nvidia GPUs. Axios reported hedge funds and trading firms are interested, and Reflection is testing a sovereign AI factory partnership with Shinsegae Group in South Korea. The open model is the front door; the factory is the business.

We don’t see that as a problem in itself. Somebody has to pay for frontier training, and selling deployment services is more honest than locking weights away. The risk is concentration of a different kind. If the Western open ecosystem consolidates around a few richly funded players tied to one chip supplier, the openness is real at the weights level and narrower everywhere else.

The competitive picture is crowded. Reflection is going up against closed labs like Anthropic and OpenAI, Chinese open models, and Western players like Mistral, Meta and Cohere. Its closest U.S. 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. That’s an apples-to-oranges win until buyers decide which capability they need.

I’m watching whether the weights ship this month as promised and whether outside testers confirm the efficiency claim. Everything else in the pitch depends on those two things.

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. 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.