Nvidia-backed startup Reflection AI debuted Beam on October 5, 2026 — its first open-weight frontier model and a direct shot at the low-cost Chinese open models that have been winning on price all year. The pitch isn't that Beam outright beats China's best open models; it's that it matches them while burning far less compute than any other American or European rival, according to TechCrunch.

Reflection says Beam scores comparably on advanced reasoning, coding and agentic benchmarks to GLM-5.2, the flagship open model Beijing-based Z.ai released in June — and is closing in on Qwen3.8-Max on coding and agent tasks. The efficiency claim is the headline: the company says Beam is three to four times more efficient than rival Western open models in token cost and inference-time compute, crediting high-compute reinforcement learning during training that teaches the model to reason in fewer steps.

On paper, Beam is a 501-billion-parameter text-only mixture-of-experts model with just 23 billion active parameters per task, trained on 23.8 trillion tokens with a 1-million-token context window — against GLM-5.2's 744 billion total and 40 billion active parameters. Weights are due later this month, so the benchmark claims are not yet independently verifiable. Founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, the Brooklyn startup has struck compute deals with SpaceX's Colossus 2 data center and cloud provider Nebius, and — per the Wall Street Journal — held talks to raise $2.5 billion at a $25 billion valuation.

Sources