After ChatGPT’s breakout on 30 November 2022, the AI world split into two camps: closed, API‑only models run by a few big labs, and open‑weight models whose code and parameters anyone can download, inspect, and run on their own servers. Open models emerged as a strategic response — meant to spur innovation, cut costs, keep data in‑house, and prevent a handful of companies from controlling all frontier intelligence.
Today, China leads the open-weight race by both pace and performance, with labs like Z.ai (GLM series), Moonshot (Kimi), and Alibaba’s Qwen pushing large, capable models into the open. In the West, Meta’s Llama family has been the main open contender, but until now there was no serious US-born open model matching Chinese reasoning and coding benchmarks at similar scale.
Reflection AI, a $25B startup by ex‑DeepMind researchers backed by Nvidia, has finally entered this arena with Beam, a 501B‑parameter “mixture‑of‑experts” model built for coding, reasoning, and AI agents. Reflection claims Beam matches China’s GLM‑5.2 on key reasoning tests while using 3–4× less compute, and it will release full weights under Apache 2.0 later this month for anyone to run or fine‑tune.
In the US landscape, Beam is the first credible open‑weight answer to China’s open models, though it still trails the very top Chinese systems like Moonshot’s Kimi K3 on raw ability. It is not yet a market‑shaking game changer, but by offering strong performance at lower cost and true local deployability, it narrows the China–US gap and gives enterprises and governments a Western “AI factory” option.
OPEN AI IS NO LONGER A CHINESE MONOPOLY — THE WEST HAS ITS FIRST REAL OPEN GUN.
Sanjay Sahay
Have a nice evening.