US Startup Reflection Unveils Beam AI Model to Challenge Chinese Giants
American startup Reflection has unveiled Beam, a 501-billion parameter open-weights AI model trained on over 10,000 Nvidia Blackwell chips to rival Chinese models.
Since the breakthrough of DeepSeek, China has dominated the open-weight language model market with Moonshot AI and Z.AI alongside tech giants like Alibaba. However, a heavily anticipated Western counter-response has finally arrived. A new American startup, backed by billions of dollars in investments from tech titans, has unveiled a model designed to rewrite the rules of the game by offering top-tier performance, efficiency, and development flexibility.
The Birth of Reflection and the Beam Model
Founded about two years ago by Misha Laskin and Yannis Antonoglou, two former Google DeepMind researchers, the American startup Reflection has raised nearly $5 billion from prominent venture capital funds such as Sequoia, private investors like former Google CEO Eric Schmidt, and the world's most valuable company, Nvidia. The startup revealed its inaugural model, named Beam, which is positioned as the Western alternative to DeepSeek.
Beam is an open-weights model featuring 501 billion parameters, with 23 billion active parameters per token. Specifically engineered for reasoning tasks, coding, and agent execution, Reflection highlights that its primary advantage lies in inference efficiency. The model delivers greater intelligence per token at a lower cost compared to its Chinese competitors, having been trained on roughly 24 trillion tokens comprised of web data and proprietary datasets.
Training Scale and Emergent Capabilities
According to the company, the training process took just four weeks but demanded an immense hardware footprint—utilizing over 10,000 Nvidia Blackwell chips. During training, the model generated over a million closed simulation environments daily with a maximum context window of 256,000 tokens. The startup notes this was one of the largest training runs ever conducted in open laboratories, driven by newly developed asynchronous reinforcement learning algorithms tailored for massive scale.
"Beam delivers greater intelligence per token at a lower cost compared to its Chinese competitors, utilizing advanced asynchronous reinforcement learning algorithms." — Reflection
Performance benchmarks released by the company show that Beam does not yet outperform leading Chinese models such as Qwen 3.8 Max, Kimi K3, or GLM-5.3. Nevertheless, it does surpass Western competitors like Nvidia's Nemotron 3 Ultra and Inkling from Mira Murati's Thinking Machines Lab. Despite lagging in certain headline benchmarks, the startup insists its model wins decisively on economic efficiency.
Remarkably, the model demonstrated emergent capabilities it was never explicitly trained for, spontaneously learning to search the web, query other models, and process image-to-text inputs despite being a purely textual model capable of processing visual information represented as text. Currently undergoing final security assessments and safety evaluations, Beam is available to a limited waitlist group and will launch broadly later this month under the Apache 2.0 license.
