By Jon Martindale
Publication Date: 2026-10-06 14:46:00
A US AI startup backed by Nvidia has revealed its new Beam AI model, claiming that it can compete with the best open-weight models in the world. Beam uses 501 billion parameters but only leverages 23 billion per task, making it several times more efficient than alternative open-weight designs from Chinese standouts developed by Alibaba and Z.ai.
Open-weight models, particularly those from China, have caught the frontier AI industry by surprise several times over the past few years. The first was the “Deepseek moment” in 2025, when Deepseek-R1 competed directly with some of the latest AI models at a fraction of the cost. The launch of Kimi-K3 in 2026 spooked them in much the same way.
But Chinese companies aren’t the only ones developing powerful open-weight models. Nvidia-backed Reflection has now launched its own model, and the company claims it’s competitive with GLM 5.2 from Z.ai and is near the performance of Qwen 3.8-Max for coding. However, Kimi K3 still remains the most capable.
The big selling point with Beam, though, is how affordable it is to run. With its 501 billion parameters, it’s not some super-lightweight model, but it uses a “mixture-of-experts” method to delegate. It distributes workloads through specialized networks within the model, meaning it rarely requires more than 23 billion parameters for any particular inquiry. That makes it around twice as efficient as GLM 5.2.
Using three to four times less compute power than alternative…
