DeepMind says Gemini 4 is in post-training, eyeing an early launch
Google's next flagship is months out, not weeks — and the summit that has dominated the week conspicuously skips the chip question.
Koray Kavukcuoglu, Google DeepMind's senior vice president, said Wednesday that Gemini 4 is "in the early phase of post-training" and that he hopes to release it "much earlier" than the end of the year. He was speaking at The Information's AI Agenda Live Summit, in his first interview since taking over the lab. The phrasing is the news: "post-training" is the stage after a model has absorbed its training corpus, where reinforcement learning, preference data and evaluation work shape what it actually does — so a model in early post-training is months from release, not weeks.
What he did not give is as telling as what he did. No benchmarks, no parameter count, no pricing, no model identifier, no date. Kavukcuoglu's claim is a direction of travel, not a launch schedule, and anyone reading it as "Gemini 4 ships this quarter" is reading past the hedge. The lab's own framing is that it wants the result out as soon as it is good, and will keep iterating after that.
The direction is the story. Google's last full flagship generation was Gemini 3, and the company has spent this year shipping point releases — Gemini 3.8 Flash, and a 3.8 TTS model that clones a voice from 30 seconds of audio — while OpenAI and Anthropic took most of the headline slots. Compressing the gap between generations is how a lab signals it intends to be in the conversation next quarter. Developers already had a hint of where this was heading: a stealth model hit the Arena this month and the people testing it said it was Gemini 4 Pro.
The US-China trade truce was extended by two months, to January 10, as Xi Jinping began his first US state visit in more than a decade. Treasury Secretary Scott Bessent announced the extension on Fox News, and the short window is the signal — many in Washington had expected six months or longer. The AI-specific part is what is not on the table: chip export controls are explicitly off the agenda, even as the two sides set up a channel for notifying each other about AI incidents serious enough to raise national security concerns.
A two-month extension is what you agree to when you want leverage next quarter. For AI, it leaves the compute question — who can buy which chips, and where Chinese models get trained — unresolved into 2027, which is precisely when Huawei and Alibaba say their next AI chips reach the market.
What to watch: whether Gemini 4 lands before January 10, and which deadline moves first.
Which shapes the AI race more over the next six months — a new frontier model, or a chip-truce deadline? Tell us in the comments.
Sources: The Information · Techmeme · NBC News · SCMP · CNBC
