Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost
TechCrunch Rebecca Bellan
Reflection just launched Beam, a big open AI model aimed at Chinese rivals. It says Beam does the same work with far less compute, which is the whole fight now.
Based on reporting by TechCrunch, Rebecca Bellan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Reflection AI is stepping into the open-model race with Beam, its first frontier model and its first big public swing at the current leaders. The Brooklyn startup says Beam matches top Chinese open models on advanced reasoning tests while using far less compute, a claim that puts it right into the middle of the fight over who gets to define the best open AI outside the closed labs.
Beam is a text-only mixture-of-experts model built for reasoning, coding, and agentic tasks. Reflection says it was trained with high-compute reinforcement learning and that it can do the work at “a fraction of the token cost and inference time compute” of rivals. The company says the model has 501 billion parameters, 23 billion of them active, and a 1 million token context window. It was pre-trained on 23.8 trillion tokens.
The pitch is not subtle. Reflection says Beam performs on par with Z.ai’s GLM-5.2 on advanced reasoning benchmarks and beats leading Western open models while using “3-4x less inference compute.” Those numbers have not been independently checked, but they are enough to sharpen the pressure on Western labs that have watched Chinese open models set the pace. Reflection is also taking aim at the usual closed-model crowd, from Anthropic and OpenAI to Western open contenders like Mistral, Meta, and Cohere.
There’s a direct U.S. comparison here too: Inkling, the open model from Mira Murati’s Thinking Machines Lab, released in July. Reflection says Beam beats Inkling on four coding tests where both companies report results, though Inkling is multimodal and Beam is not. That matters, because Beam is being sold as a workhorse for enterprises, the public sector, and developers, not a flashy demo model.
The company has the funding and the chips to make the claim feel less like marketing and more like a plan. Founded in 2024 by two former Google DeepMind researchers, Reflection has raised roughly $4.7 billion, according to PitchBook, and its last round valued it at a $25 billion pre-money valuation. This summer it signed more than $7 billion in deals with SpaceX and Nebius to get access to Nvidia’s GB300 chips through 2029. Reflection says Beam’s weights and full technical details will arrive this month, along with distribution through hyperscalers and neoclouds and integrations across open source libraries at launch.
My take — AI-written commentary, not fact-checked reporting
This is the part where open AI stops being a philosophy seminar and becomes a spreadsheet war. Reflection isn’t selling purity; it’s selling cheaper inference, bigger context, and enough compute muscle to make sovereign AI sound less like a slogan and more like procurement. That’s a lot more credible than the usual open-model victory lap, and also a reminder that “open” still needs very expensive plumbing.
Read more about this at: TechCrunch
