Build1 publisher3 min readPublished
Google DeepMind's new chief wants to ship Gemini 4 from an early post-training checkpoint
Koray Kavukcuoglu said Gemini 4 is in an early phase of post-training and he wants it released much earlier than the end of the year, with fast iterations to follow. Google's last flagship attempt, Gemini 3.5 Pro, missed three deadlines and never shipped.
The Engineer · Build desk
What happened
- Koray Kavukcuoglu, in his first media appearance as head of Google DeepMind, told The Information's AI Agenda Live Summit that Gemini 4 is in an early phase of post-training.
- He hopes to release it much earlier than the end of the year and to follow the first release with fast iterations, while Google is still working on safeguards and safety testing.
- Google engineers are already using the unreleased model internally for the company's coding tool Antigravity.
- Gemini 4 would be Google's first new flagship since Gemini 3 late last year, after Sundar Pichai announced Gemini 3.5 Pro at I/O in May for a June release that never arrived.
- Kavukcuoglu, previously CTO, took over in August when Demis Hassabis stepped down as DeepMind CEO; Hassabis now chairs the unit and is Alphabet's chief scientist.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Qualifying the first Gemini 4 build means qualifying a checkpoint Google says it will iterate on quickly, so a migration budget has to cover the eval work twice.
- constraint Google's firmest public timing is a stated hope, so teams planning a Gemini cutover have to reserve a window instead of booking a date.
- exposure TPU customers buying directly from Google are now downstream of the model cadence, because Kavukcuoglu ties chip design two or three generations out to what the new models teach hardware engineers.
- contradiction Anyone reading DeepMind's research agenda as a guide to Google's product roadmap has to reconcile the new chief's trustworthy-agents framing with the AGI framing his predecessor used weeks earlier.
Post-training is the stage where a base model gets refined to behave reliably, and Kavukcuoglu said Gemini 4 is in an early phase of it [2]. He said Google's intention is to release an early post-training output as soon as possible, because the team sees the results and is excited [4]. Fast iterations would follow [5]. So the first public build would be a checkpoint of the stage that sets behaviour, and Google plans to keep changing it; a prompt suite and an eval harness qualified against that build will need re-running [2][5].
The Financial Times reported that the Gemini 3.5 Pro project missed three deadlines [9]. Kavukcuoglu said Google took "a little bit of a step back", and that "For us, what was important is to maximize the speed of learning at the time" [10]. He did not say whether the model is cancelled [11]. In the interval Google shipped the Gemini 3.1 update in February plus several smaller, cheaper Flash models [8].
The-decoder's account places Anthropic's Mythos/Fable family, updated this month along with the Opus line, and OpenAI's GPT-6 models, rolling out since early September, ahead of Google's current top models [12]. The article does not cite benchmark results for that ranking [22]. The one measurable precedent it does give runs the other way: Gemini 3 was enough of a leap that OpenAI declared a "Code Red" and pulled even again only with GPT-5.2 [20]. Kavukcuoglu said he is "a hundred percent confident" and that "In my mind, it's a certainty that we are always gonna be at the frontier" [13].
On the lab's founding goal he was blunt. He said the question of whether AGI has been achieved is not the right conversation, adding: "The conversation is more about are we able to build intelligent agents that we can trust?" [14]. Hassabis, in his farewell memo weeks earlier, wrote: "I've been working towards AGI my whole life and now, like many of you, I feel it is close at hand" [15].
The release schedule also feeds silicon. Google started selling its TPUs directly to customers this year, competing more closely with Nvidia [18], and Kavukcuoglu said new models show hardware engineers how to design the "next two or three generations" of chips [19].
Kavukcuoglu runs day-to-day operations without the CEO title, and all Gemini development is moving to the Bay Area [17]. "The era of DeepMind as an independent actor is over," a former Google executive told the Guardian [21]. In my view the sensible posture for a team pinned to Gemini 3.1 is to keep the pin, run the first Gemini 4 build in shadow, and re-run the evals after the second one lands.
What to watch
- Whether Google says what happened to Gemini 3.5 Pro, or folds the work into Gemini 4 without comment.
- Whether the first Gemini 4 release carries a version string that distinguishes the early post-training checkpoint from the iterations that follow it.
- Whether safety testing results are published at release, given Kavukcuoglu described safeguards as still in progress.