When, if ever, do you think we will have an open-source Gemini 3 Pro?

Posted by Unusual_Guidance2095@reddit | LocalLLaMA | View on Reddit | 25 comments

I think Gemini 3 Pro is the most impressive model I have used to date. I’m not interested in half-assed runnable-but-nowhere-near-as-good local models; I want a model with all the capabilities Gemini 3 Pro has. I don’t care if I can’t run it locally without being rich, I just want the weights to be open sourced so I can have the flexibility of switching cloud providers, no secret fine-tuning, extra censorship, etc., all the other benefits of open-weight and maybe if I’m rich enough one day… . I don’t want my customers, if I use this for my business, receiving BS results even if it costs more and I don’t understand how other businesses are okay with running a 8b-q4 model without feeling like they’re ripping off their customers or how they trust such models with anything in their business when even the SOTA ones suffer from certain problems but these smaller ones suffer like 1000x. For personal use, I want the ability to be used in Antigravity, full computer use ability, and like nano banana pro, image creation and of course, state-of-the-art reasoning, multimodality, etc. The best model I can see Kimi K2 comes no where even close: beginning with just the multimodality stuff. We keep saying that open-source is 9 months behind but I mean single-model multimodality at the very least seems to be lacking years behind, reasoning now seems to be lacking at least a year behind, vision seems to be significantly more, and open-source toolkits built around these models like VSCode forks even longer, like I can’t name one open source Antigravity, Cursor competitor. Do you think this is something I should hold my breath for or just accept that I’ll always have to use a closed source model to get Gemini 3 Pro level results?

[-]

jacek2023@reddit

Please go back to 2023 and ask random person is it possible to have model like ChatGPT running locally at home computer.

[-]

ForsookComparison@reddit

Chatgpt4 at home was always endgame.

Now we have it. It runs on modest gaming rigs, not even A.I. enthusiast rigs.

And nobody is doing cool stuff with it.

[-]

Everlier@reddit

So the answer is: "sooner than predicted by sceptics but later than hoped by optimists"

[-]

Ambitious-Profit855@reddit

The answer is "by the time you can have Gemini 3 Pro at home you want Gemini 5 Pro and not some "half-assed runnable-but-nowhere-near-as-good local models"".

[-]

Unusual_Guidance2095@reddit (OP)

Yeah I understand, there has definitely been fast progress in the last few years no doubt, I am already really impressed with Kimi, Deepseek, etc. and I doubted a few years ago a bit that I would ever have access to some of this stuff. Trying the initial open source models at the time which barely produced coherent sentences to now is definitely a massive leap. But I guess I’m just being human: never satisfied, thought Kimi K2 was the be all end all but now wanting more because of Gemini’s massive leaps, etc. But I think honestly some people would have probably known that eventually we would get something comparable to ChatGPTat home, I think I did, I think I was more sure something like ChatGPT-3 could become open-source than I am now about Gemini-3 Pro since everything has changed so much: Chatbots are no longer a novelty everyone is working on, and it seems to me, development has started concentrating towards a few companies with everyone else giving up, those companies are increasingly releasing closed source models (like I can’t remember the last Gemma) etc.

[-]

MitsotakiShogun@reddit

Or go back to 2020 and ask an AI researcher if zero-shot AI with over-random accuracy (even on the most narrow domain) is possible in this decade.

segmond@reddit

I want cookies, chocolate chip cookies with exactly 17 chips, 13 visible and 4 inside the cookies.

I want cookies.

Agitated_Space_672@reddit

In some tasks I think ring and kimi-k2-thinking can beat Gemini 3 already.

HarambeTenSei@reddit

I mean most of the utility comes from google searching and smartly handling the RAG context. I highly doubt that the benefits actually come from the raw model itself.

stoppableDissolution@reddit

...and the model itself, while good at long context, has a bunch of short attention issues

Terminator857@reddit

Interesting, any examples?

I'm swiping it away, but after 20-30 messages in the conversation it might start ignoring your last question and reply to second/third to last instead

Thanks, I noticed that too. I thought it was on odd quirk with my questions.

Pvt_Twinkietoes@reddit

And that should be the way we work with these models.

lisploli@reddit

You will never have "the most impressive" since there is no one telling you advertisement phrases for your own setup.
Software progression makes up for the lack of hardware after some time, but the frontier obviously doesn't wait.
Better ask if it's good enough for the task.

aizvo@reddit

Qwen3-VL can already give you 90% of frontier experience on 24GB VRAM.

I'm a bit surprised people like Gemini so much. For me claude does a better job, but I'm mostly doing coding.

Macestudios32@reddit

That phase is over... Wow can analyze images! Qwen3-vl, gemma3 You can talk to him and watch qwen3 omni videos. I'm quite happy with the evolution, I'm more concerned about censorship, hw control and legislation than the progress itself. I feel lucky for the times we live in and you can have yourself at home

Longjumpinghy@reddit

6 months

NickWebster@reddit

The capability of open weight models typically lags that of the (publicly accessible) SOTA from the big tech companies by about 9 months. The tooling around them is probably a little closer though.

reginakinhi@reddit

Are we in that stage of the eternal cycle again already?

I don't know what you aim to hear here, obviously open source models can't quite compete with a brand new preview frontier models from the company that pioneered TPUs. As always, we will slowly see open source catch up in different aspects over the course of 9 months to a year. The pure LLM performance will most likely come close or be exceeded in that time, with multimodality being a less often covered topic. I presume Qwen 3.5 or Qwen 4 VL will be closing the gap there (or GLM-5V, for all I know).

Image generation has always been a bit slower, since it's just less 'worth it', but I presume such needs will also quickly be covered by the next iteration of Qwen-Image or something else entirely.

Either way, incoherent rambles such as this post serve no purpose. Everyone here will be well aware that Open Models aren't instantly on the level of brand new proprietary ones - shocker, I know - and complaining about it constantly does not help anyone.