Reflection AI, a two-year-old Brooklyn startup backed by Nvidia, unveiled Beam on Monday — its first frontier open-weight AI model. With 501 billion parameters and a bold pitch as the Western answer to DeepSeek and Qwen, it's one of the biggest open-weight launches of the year.
The specs
Beam is a text-only mixture-of-experts model. The numbers: 501 billion total parameters, but only 23 billion active for any given task — which is how it stays fast and cheap to run. It was pretrained on 23.8 trillion tokens and carries 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 trained Beam with high-compute reinforcement learning, targeting reasoning, coding, and agentic tasks — the “workhorse” jobs enterprises actually pay for.
The big claim
Reflection says Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks and outperforms today's leading Western open models while using 3-4x less inference compute. Its own benchmarks also show Beam beating Inkling — the open model from Mira Murati's Thinking Machines Lab — on four coding tests (Inkling is multimodal; Beam is text-only). One big caveat: none of this has been independently verified. Until the weights ship and third parties test them, treat every benchmark as company-reported.
Why it matters
Open-weight AI is where the real price war is happening. Chinese labs have owned it — DeepSeek, Qwen, Kimi are cheaper, customizable, and nearly as capable as closed models. Now US tech firms want back in. Reflection's angle is “AI factories”: selling enterprises and sovereign nations their own local AI systems built on Reflection's models, trained on proprietary data. It's already testing a sovereign AI-factory partnership with Shinsegae Group in South Korea. Nvidia's Jensen Huang has been pushing this AI-factory vision for years — and Nvidia is one of Reflection's backers.
The money behind it
Reflection was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. It has raised roughly $4.7 billion from Nvidia, Sequoia Capital, and Lightspeed Venture Partners (per PitchBook), at a $25 billion pre-money valuation. This summer it locked up more than $7 billion in compute deals with SpaceX and Nebius for Nvidia GB300 chips through 2029.
What to watch
Reflection says Beam's weights and full technical details arrive this month, distributed through hyperscalers, neoclouds, and open-source libraries. Until then, the “cheaper than DeepSeek” claim is just a claim. But if the weights ship and the benchmarks hold, the open-model race just got a serious American contender.
Comments
Post a Comment