Reflection AI releases Beam, a 501-billion-parameter open model
The lab pitches it against Chinese open models and sells "AI factories" to enterprises and governments.
Reflection AI released Beam on October 5, an open-weight model it is aiming squarely at China's leading open models. Beam is text-only, a mixture-of-experts system built for reasoning, coding and agentic work, TechCrunch reported.
The numbers are big. Beam has 501 billion total parameters, 23 billion of them active at any moment, and it was trained on 23.8 trillion tokens. Its context window runs to 1 million tokens, enough to hold a large codebase or a stack of documents in one prompt. Only a slice of those 501 billion parameters fires on a given request. That is how something this large can still be comparatively cheap to run.
Reflection says Beam matches the best Chinese open models on advanced reasoning benchmarks, landing on par with Z.ai's GLM-5.2 and ahead of Western open models. The company also claims Beam needs three to four times less inference compute than rivals, a pitch aimed straight at buyers who watch the cost of serving a model.
Open weights plus lower running costs is the whole argument. An open-weight model can be downloaded and run on a buyer's own machines instead of called over an API, and Reflection is pointing Beam and the models after it at enterprises and sovereign nations that want to keep AI in-house. The bet: the next wave of buyers will be institutions that cannot, or will not, hand their data to someone else's cloud.
The pitch is to build “AI factories,” a product that would let institutions build their own customized, local AI system by training Reflection's AI models on their own proprietary data. Reflection calls Beam a workhorse for enterprises, the public sector and developers, and it is already testing the factory idea with South Korea's Shinsegae Group.
The money is there to chase that market. TechCrunch reported Reflection has raised about $4.7 billion, at a $25 billion pre-money valuation in its last round, with Nvidia, Sequoia Capital and Lightspeed Venture Partners among the investors. It also holds compute deals worth more than $7 billion with SpaceX and Nebius for Nvidia GB300 chips through 2029.
Sources
- Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost · TechCrunch, AI