Nvidia-Backed Startup Launches Efficient AI Model to Challenge Chinese Rivals

An Nvidia-backed AI startup has introduced Beam, a new open-weight model designed to rival leading alternatives from Chinese labs. Beam boasts 501 billion parameters but operates with just 23 billion per task, making it significantly more efficient than models from Alibaba and Z.ai. Open-weight models have previously disrupted the AI industry, with Deepseek-R1 in 2025 and Kimi-K3 in 2026 standing out for their cost-efficiency and performance.

Reflection, the startup behind Beam, claims its model is competitive with GLM 5.2 from Z.ai and nearly matches Qwen 3.8-Max for coding tasks. However, Kimi K3 remains the most capable model overall. Beam's efficiency stems from its "mixture-of-experts" method, which delegates workloads through specialized networks, reducing compute power needs by three to four times compared to other open-weight models.

The affordability and speed of Beam make it an attractive option for developers looking to create custom models. Open-weight designs allow users to run models on any hardware, providing flexibility and cost savings. AI benchmark platform Artificial Analysis, which had early access to Beam, reports that the model is on track to be one of the most efficient at its intelligence level.

This development follows a public letter signed by major US tech companies, including Nvidia, urging the US government to support open-weight model development. The White House later shifted its stance, indicating a potential change in policy to address international competition. Reflection, founded in 2024 by former Google DeepMind researchers, has raised over $4 billion, with $800 million coming from Nvidia.