A New York AI startup founded by two former Google DeepMind researchers has released its archetypal important example — and is pitching it as a homegrown answer to China's increasingly mighty open AI systems.
Reflection AI on Monday unveiled Beam, an "open-weight" example designed for coding, reasoning, and AI-agent tasks. The business says it rivals several of China's foremost open models during using far small computing power.
Here's what to know.
What is Beam?
Beam is Reflection AI's archetypal frontier open-weight model. Open-weight method developers can download the model's parameters, run them on their own hardware, and modify or fine-tune them for particular use cases.
That is an substitute to using models from companies specified as OpenAI and Anthropic, which are mostly accessed via their own Application Programming Interfaces (APIs).
Reflection says Beam, a text-only model, has 501 milliard total parameters, but activates lone 23 milliard at a period — a scheme intended to create it cheaper and faster to operate. The example can grip coding, multi-step reasoning, and another agentic work.
For now, Beam is in final safety evaluation and accessible through a waitlist. Reflection says it volition publish the model's weights, a specialized report, and developer tools afterward this month.
Who is rearward Beam?
Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, the two erstwhile DeepMind researchers.
"The lone way to own intellect is, by definition, if it's open — since if it's open, you can run it on your stuff, you can own it, you can authority it, you can customize it," Laskin, the CEO, stated in a podcast released on Monday.
The startup has raised billions of dollars and is backed by investors including Nvidia, Sequoia, and Citigroup. In June, it confirmed a backing circular at a $25 milliard pre-money valuation.
How does Beam compare?
Reflection stated Beam is rivalrous alongside Z.ai's GLM-5.2 on coding and agentic tasks, and is approaching Alibaba's Qwen 3.8-Max. It stated that Moonshot AI's Kimi K3 remains onward on raw capability.
The startup is pitching itself on efficiency. It stated Beam can accomplish results comparable to GLM-5.2 on advanced reasoning benchmarks during using three to four times small conclusion compute — the handling required for a example to answer a prompt.
How are tech and endeavor leaders reacting?
Beam's initiate is being welcomed as a possibly crucial US reply to China's open-model lead. It arrives in a crowded site that includes open models from Meta, Mistral, Nvidia, and Thinking Machines Lab.
Beam additionally faces apparent competition from mighty models released in China by companies including DeepSeek, Alibaba, Z.ai, and Moonshot AI.
The Wall Street Journal reported in March that several investors had described Reflection as the "DeepSeek of the West." That comparison to DeepSeek has rotate into the shorthand used in many X posts surrounding Beam's buzzy launch.
On Monday, Nvidia AI's authoritative X document congratulated the business on the launch. Investors, including Deedy Das and
Shaun Maguire, additionally expressed their enthusiasm for the initiate on X.
Read next
Shuby is a elder newsman at Business Insider's Singapore bureau, anywhere she writes concerning tech, alongside a concentration on AI training and vibe-coding startups.She earlier interned for Bloomberg News and CNBC. She studied communications and endeavor at the National University of Singapore. In 2026, she won the Singapore Press Club Tech Journalism Award for her activity on the hidden individual disbursal of Meta's chatbot training program and an exposé on the dreary market for AI gig-work accounts.Have a tip? Contact her via email at [email protected] or on safe messaging app Signal at @shuby.85. We can keep sources anonymous.
