Google Should Open-Weight Its Best AI Model Yet

Google is facing pressure to release an open-weight version of its Gemini model, which would be a significant move in the AI market. The company's research organization has been losing talent, with key figures such as Noam Shazeer and John Jumper leaving for other firms like OpenAI and Anthropic.

The departure of these researchers comes at a time when Chinese labs are dominating the open-weight layer of the AI market, with models like Kimi K3 and DeepSeek v4 being widely used. In contrast, Google's Gemini model is not yet available in an open-weight version, which has led to concerns about concentration of power and lack of control over data and infrastructure.

Nvidia has already released a 550-billion-parameter mixture-of-experts model called Nemotron 3 Ultra, which is the first American-origin open-weight model to break past the ceiling set by Google's Gemini. This development suggests that incentives are driving the creation of open-weight models, rather than just research labs.

Releasing an open-weight version of Gemini would allow Google to compete directly with OpenAI and Anthropic on a different axis: price. By making its best reasoning available for free, Google could put pressure on these firms to lower their prices, which has been demonstrated by the release of Meta's Llama models.

Additionally, an open-weight Gemini model would be the first mainstream, lab-built American alternative to Chinese-origin models, providing a much-needed counterbalance to the current market dynamics.