Google has started Pre-Training Gemini 4

https://twitter.com/OfficialLoganK/status/2079594867161022817

by EvanZhouDev • 1 day ago

3 4 1 day ago

Every time there's any news or rumors about a new Gemini model, I cross my fingers and hope that it'll be the one that finally catches up to other big AI coding models. It hasn't happened yet, but I'm hoping this time it will... :')

Rzor• 1 day ago

These days I'm more interested in Gemma models. Can't wait for Gemma 5. Google crushed it.

Topfi• 15 hours ago

That's insane. Pre-training a frontier model is a multiple 100 Million USD moonshot (pun somewhat intended). You want to get as much out of one as you possible can, squeeze that stone till there is nothing more to give via post-training, if they not only start another run with Gemini 3.5 Pro still not being released and feel the need to announce that publicly, that's easily half a Billion, not accounting for man hours and ancillary costs, that they loose on the 3 Pro base right there.

Purely encoded information wise, there are some things in the Gemini 3 Pro series that (with the exception of Inkling) no other model could accurately answer. But if this is a 2, 5 or worst-case 10T pre-train they cannot get to reliably tool call and output in a somewhat token efficient manner, that is an indictment. Hopeful to be wrong, we need competition in all forms, but along with 3.6 Flash (which in task completion pricing and full task run speed ends up more expensive vs GPT-5.6 Sol at worse results in my evals) I am frankly astounded what is happening here, especially as their efforts with Gemma 4 were incredibly impressive.

This could be a true gamble by them, putting all their chips into hopefully something that outperforms the frontier including on tool calling and is token efficient, which is very risky indeed and also somewhat desperate from a lab that not long ago (Gemini 2.5 Pro) was ahead across most metrics. I suspected a lack in tool calling data to train on initially during the Gemini 3 Pro launch, but the way they hyped that up was still utterly ridiculous and after months of Antigravity including that providing Opus data, this should have been resolved, so seems there are deeper issues here.

If they pull this off, then they only need to address all the account issues, deprecation of popular CLIs, all the bugs in Antigravity, the mess that is their documentation, all the overall confusion with workspaces accounts, the major difference in experience between Geminis main website and AIStudio (the former having certain bugs for months that the latter never had), etc., which are all no small feat in themself, but that's for then to worry about.

lern_too_spel• 5 hours ago

The other frontier labs are of course pre-training their next major versions, but they have no reason to pre-announce because they aren't so far behind.

Google has started Pre-Training Gemini 4

Related Stories