Reflection AI, a Brooklyn-based startup, unveiled Beam, its first open-weight AI model, on October 5. Beam features 501 billion parameters with 23 billion active ones and was pre-trained on 23.8 trillion tokens. The model offers a 1 million token context window and aims to rival leading Chinese AI models like DeepSeek, Qwen, and Z.ai at a significantly lower compute cost, according to techcrunch.com.
The company detailed Beam in a blog post, describing it as a text-only mixture-of-experts model trained with high-compute reinforcement learning. Beam is designed to excel in reasoning, coding, and agentic tasks while using three to four times less inference compute than competitors. Reflection AI claims Beam matches the performance of Z.ai’s GLM-5.2 on advanced reasoning benchmarks and surpasses current Western open models, positioning it as a practical solution for enterprises, the public sector, and developers, per techcrunch.com.
Beam’s launch enters a competitive AI landscape dominated by closed labs such as Anthropic and OpenAI and popular open models from Chinese developers. With 501 billion parameters, Beam is smaller than Z.ai’s GLM-5.2, which has 744 billion total parameters, but claims comparable performance at reduced computational expense. This cost efficiency could influence adoption in sectors requiring scalable AI solutions, reflecting a growing demand for powerful yet affordable AI models, according to techcrunch.com.
Reflection AI’s Beam model represents a strategic effort to provide a Western alternative to Chinese open-weight AI models while lowering operational costs. The company’s announcement follows reports from Axios and includes detailed performance claims in its blog. Beam’s development timeline spans two years, emphasizing Reflection’s focus on balancing scale, performance, and cost in AI model design, as reported by techcrunch.com.