Google DeepMind's Kavukcuoglu: Gemini 4 in early post-training, launch expected well before year-end

Google DeepMind's Koray Kavukcuoglu said on Wednesday that Gemini 4, Google's next flagship AI model, has entered the early phase of post-training and that the company hopes to release it "much earlier" than the end of the year.

Kavukcuoglu, a senior vice president at the lab, spoke at The Information's AI Summit in his first interview since taking charge of DeepMind. The publication reported his remarks on Wednesday.

What the post-training phase means

Post-training begins once a base model's large-scale pre-training run is finished. Teams use human feedback, reinforcement learning and specialised data to sharpen a model's reasoning, coding, tool calling, safety and instruction following.

"Our intention is to release an early post-training version as soon as possible, as we have seen results and are very excited," Kavukcuoglu said, adding that Google will "continue with rapid iterations".

The company has not disclosed a release date, a parameter count, training costs or benchmark results, and it is not yet clear how wide the first rollout will be.

Pre-training started in July

In July, alongside the launch of Gemini 3.6 Flash, Google said it had begun "our most ambitious pre-training run yet, for Gemini 4". A day later, Alphabet CEO Sundar Pichai repeated the line on the company's second-quarter earnings call, saying Google was excited by the progress it was seeing at the frontier.

Why the timing matters

Google is widely seen as having lost ground to OpenAI and Anthropic in the race to build frontier models, and Gemini 4 is its main answer. A strong launch would support the argument that its heavy spending on AI research and infrastructure is producing competitive products; a weak debut would reinforce doubts.

Its advantage is distribution: a successful model can be pushed into Search, Google Cloud, Workspace and Android, reaching consumers, developers and enterprises at once — reach that standalone labs selling subscriptions and API access cannot match. No date is set, so for investors the effect is still anticipation rather than a repricing; the first performance data will be the real test.