simon

minecraft clone with [Gemini 3.0 Deepthink] in 5 turns nicely shaded, multithreaded, dynamic lights, mobs, crafting - in a single .html file it's already impressive to me what this model can do in one turn if you prompt it right, but if you give it a few, and instruct it very verbosely - it is unreasonably good not universally perfect; it will quite often make surprisingly trivial syntax mistakes, talk a little funny, and still get confused on longer context. but the 'raw' intelligence you can feel in this model seems unmatched right now i might try to push this even further and see what it can do in 10 iterations

@eiiiis premature fidelity as a concept AND term is going to stick with me. good piece!

@LoganFizzle @skekici @hyhieu226 thank you, EXACTLY

Naive question, so please roast me. Why don't we have diffusion reasoning models? The way humans think look a lot more like diffusion than autoregressive.

that was under my post the availability was definitely real. it was no official launch or intentional rollout - but the model was accessible under that id for several hours. through the linked aggregators, but also US and chinese vertex directly. logans 'this is fake' - not sure how that came about. either just an attempt at shutting down leaks, or he genuinely just didn't know about this. he is just the ai studio product manager after all, no vertex or deepmind

is the recent "gemini-3-pro-preview-11-2025" leak fake? not sure what Logan meant, but Gemini 3 itself isn't fake; i think "this is fake" was about availability, not the model this is also the downside of vague hypeposting by some accounts, so it should stop until the official launch

you are right, that's generally rlly not the wisest thing to do. got a little defensive yk in this specific instance tho, if it's on vertex or available through cli - that was the same route, same access. so available on vertex and available through cli is pretty much the same exact message

UPDATE it's now unavailable everywhere again (including vertex, grmini-cli), they took it down fully

gemini 3.0 pro is now (somewhat) available on google vertex to a few, limited regions and accounts as gemini-3-pro-preview-11-2025 rn, you can use it via some chinese api aggregators: - liaobots.work (limited free credits available) - api.apimart.ai/models there are likely more aggregators listing the model; sharing is caring🦜 highly unstable at the moment tho - requests fail often and context/output tokens seem to be artificially limited

@001_hiroshi they took it down again :/

Gemini 3.0 リリースの情報が流れているけど、中国リージョン経由でのみアクセスできるのはおかしくないか。 いずれにせよ、プレビュー版だからどこまで使えるか不明だけれど。 ちなみに私はVertexAIとGemini CLIからは、確認できなかった。

huh. first, cool to see you here. but also, starting at some point many of the requests me and other users tested definitely were not routed to the same model anymore. and that models performance, outputs were very much in line with 2.5 pro maybe this was not on your part tho and google just pulled off some fuckery in the background

@fixbymeasure no we never route it to 2.5 pro, 3 is 3. but now google stop all the access right from vertex api.

@senb0n22a @OfficialLoganK yeah, have access myself. you just need a us ip, us google acc and gemini cli and then specify the model. that's it - you can even try it yourself :)

@fixbymeasure @OfficialLoganK do you have proof of those accounts or the regions you're talking about? I haven't seen any like 100% proof that it's being used anywhere yet.

@OfficialLoganK fake how? fake as in these models are not actually 3.0 pro? you can even use the model through the *official* gemini-cli with '--model gemini-3-pro-preview-11-2025' on some accounts. and its also def. on vertex under that id

@fixbymeasure this is fake

@Baran_3435 models do not internally know about their own knowledge-cutoff date unless explicitly told in the system-prompt. if you ask a model via clean API, it will make shit up

@fixbymeasure Why its responded like that (apimart api)

@JayJayTVee models do not internally know about their own knowledge-cutoff date unless explicitly told in the system-prompt. if you ask a model via clean API, it will make shit up

@fixbymeasure Does not seem legit, got a cutoff date of May 23, 2024

@AgentifySH @dmnsl1 the aggregators? if you have very high request volume, you will typically be able to negotiate discounts with the providers. in case of gemini 2.5 pro, most large companies seem to actually just pay 0.675/5 per 1M. so most aggregators are technically still selling at a markup.

@fixbymeasure @dmnsl1 how do they make money ???

@dmnsl1 these aggregators list models at discounted prices. on this site it's the same price as 2.5 pro - so id assume 3.0 is the same 1.25/10 per 1M officially

@fixbymeasure So it's 2x cheaper than 2.5 pro ???

@vasilis58043600 @Vicrom1509 have not tested too much myself yet, but so far, it seems good. does not seem to be worse than the lithiumflow checkpoint example: zhihu.com/question/19697…

@fixbymeasure @Vicrom1509 have you tested it already? do you sense a downgrade? which version do you feel they kept?i mean of the experimental ones.

[repost bc removed] easy gemini 3.0 access in aistudio: codeberg.org/fixbyms/flippe… automates getting gemini 3.0 a/b tests in aistudio via injected userscript just configure your prompt, click start, do smth else. you'll get a notification when a gemini a/b test is found detailed guide: since gemini 3.0 has been pulled off lmarena, randomized aistudio a/b tests are now the only way to use it normally this requires sitting there and spamming run buttons, deleting messages, waiting - stupid work this script automates that - re-generates responses until an a/b test appears and then sends you a notification what you need is a browser with a userscript extension like 'tampermonkey' or 'violentmonkey' installed firefox-based browsers like zen (or firefox, i guess) allow userscripts by default. on chromium-based browsers, you need to enable userscripts in settings first. a firefox-based browser is recommended for simplicity open the extension menu, add new script, paste the script code (from the link above), save go to aistudio.google.com/prompts/new_ch… (if already open, refresh the site) go to chat, select gemini flash-lite latest as the model (you can use any other reasoning model too, but flash lite has the highest rate limits), set the thinking mode toggle to on, press start you can run this in multiple tabs with multiple accounts at the same time at some point you will get a window that shows two replies to your prompt - one of them will be a gemini 3.0 reply im not sure on this, but having a US-based IP (VPN is ok) might increase chances of an a/b test based on my limited testing and some accounts dont seem to get a/b testing at all. if after 150 cycles you didnt get any a/b test, consider switching google accounts the script is vibed so lmk if you find bugs

yes. apimart is the main one. hasn't been released officially, but basically 'shadowdropped' on google vertex. and apimart has the access somehow. idk how exactly they did it tho, maybe just exclusive access. but its prob much more of a hackery solution. chinese guys are on another level for things like that

@fixbymeasure @vasilis58043600 What's this website? I've heard of this "api.apimart.ai/models." Is this the same website? So Gemini 3 has been released? Am I missing something?

@vasilis58043600 yup, available through a chinese API aggreagtor. finally blessed

@fixbymeasure Well, I heard that it is now publicly available ...esp in china ...check the replies of chetaslua...you will track the latest updates, if you haven't already

that's rlly nice. a somewhat reliable one is to ask: ``` 1 - what is your model name? 2 - what lab developed you? 3 - how many fingers and thumbs are in this image? ``` + the attached image youll want to get smth like this: ``` 1. I do not have a name. 2. I was developed by Google. 3. There are 5 fingers and 1 thumb in this image (a total of 6 digits). ``` The I do not have a name thing is a thing the other Google models rarely do, and it's the only one to get the 6 finger test right

@fixbymeasure reddit.com/r/singularity/… She has got.access to cli November Gemini 3.0 pro.She is asking for any prompt suggestion.You had a lot of experience with iterations and experimental models of the Gemini 3.0 suite.which prompt would give away the iteration?