TECHNOLOGY · AI · CYBERSECURITY · THE SYSTEMS SHAPING BUSINESS
Reflection AI, the Brooklyn startup that has raised billions to build a U.S. alternative to Chinese open models, unveiled Beam, its first frontier open-weight model, TechCrunch reported Monday. The company says Beam performs on par with Z.ai's GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute, a claim that has not been independently verified.
What Happened
Reflection announced Beam in a blog post Monday, confirming weekend reporting by Axios that a launch was close. Beam is a text-only mixture-of-experts model with 501 billion total parameters, of which 23 billion are active for any given token. It was pretrained on 23.8 trillion tokens and supports a context window of 1 million tokens.
The company's headline comparison is with GLM-5.2, which TechCrunch said has roughly 744 billion total parameters and 40 billion active. Reflection also published results against Thinking Machines Lab's Inkling, saying Beam scored higher on the four coding tests where both companies report numbers. Inkling, however, handles images as well as text, while Beam is text-only.
RelatedSiliconANGLE reported that Reflection pretrained the base model on a cluster of 6,144 graphics cards and used 10,000 Nvidia GB300 chips for the reinforcement learning phase. It also noted that Beam still trails the top closed frontier models.
Reflection is pitching Beam as a "workhorse model" for enterprises, the public sector and developers. It said it will publish the weights and full technical details this month, distributed through hyperscalers and neoclouds. Until then, the model is available through early access.
Why It Matters
Open-weight models from Chinese developers, including DeepSeek, Alibaba's Qwen and Z.ai's GLM family, have become common starting points for companies that want to run and fine-tune models on their own infrastructure. Reflection has positioned itself as a Western supplier for buyers, especially governments, that are wary of building on Chinese systems.
Compute efficiency is central to that pitch. If Beam delivers GLM-5.2-level reasoning with fewer active parameters, it would be cheaper to serve per query, which matters most for the high-volume, routine work Reflection says the model is designed for. The coding comparison with Inkling is narrower: it covers four tests and pits a text-only model against a multimodal one.
The company has the money to keep going. Founded in 2024 by two former Google DeepMind researchers, Misha Laskin and Ioannis Antonoglou, Reflection has raised about $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners, according to PitchBook data cited by TechCrunch. Its most recent round valued it at $25 billion before the new money.
Compute Deals and AI Factories
This summer, Reflection signed deals worth more than $7 billion combined with SpaceX and Nebius for access to Nvidia GB300 chips through 2029, TechCrunch reported.
Its commercial plan goes beyond releasing weights. Reflection wants to build sovereign "AI factories," in which institutions train its models on their own data on local infrastructure, and it is already testing the concept with South Korea's Shinsegae Group. Nvidia CEO Jensen Huang has long promoted the AI factory idea, which also drives demand for his company's chips.
That puts Reflection in competition not only with Chinese labs but also with Anthropic, OpenAI and other Western model makers such as Mistral, Meta and Cohere, several of which are courting the same government and enterprise buyers.
What to Watch
The release of the weights and technical report later this month will let independent researchers test Reflection's benchmark and efficiency claims on their own hardware. Other markers are which cloud providers carry the model at launch, what license terms Reflection attaches to the weights, and whether the Shinsegae pilot turns into a paying sovereign deployment that other national or corporate buyers follow.
