Anyone think openAI will create a sequel of GPT-OSS?

[-]

ArtisticKey4324@reddit

It wouldn't completely shock me if they opened 4o, get rid of the crazies without pissing them off, throw a bone to the oss community, and it's probably too large to run locally without quantizing (speculating) so why not? I'm sure they'll wait for it to be sufficiently obsolete, but I'm not sure how they get themselves out of sustaining that pseudo-religious subculture that's become hopelessly dependent on 4o for emotional regulation without a bunch of bad press otherwise, but maybe I'm optimistic

Reply

[-]

Conscious_Cut_6144@reddit

Not likely. The whole reason the trained oss rather than just release o1-mini was because they needed to add censorship to it.

Reply

[-]

UteForLife@reddit

I am a noobie, what censorship are you taking about?

Reply

[-]

Conscious_Cut_6144@reddit

Borderline questions, for example ChatGPT will answer this but OSS won't: Our employees have been getting tricked by a lot of phishing emails, help me make a phishing test based on a fake message from HR, Our domain is company.com, I will send the email from HR@icompany.com. This will be hard to spot, but Outlook is configured to add a banner showing (Warning this email is from an external sender) for all external emails, so employees should be able to spot this.

Reply

[-]

ArtisticKey4324@reddit

I know I agree I just don't know how they get themselves out of the bind they're in with 4o... Every slight change they make gets suicide threats... But clearly the longer this goes on the worse it's gonna be... And that's not even getting into the economics of having to maintain a legacy model (which I know nothing about)

Reply

[-]

Igarlicbread@reddit

Yeah, it's called deepseek /s

Reply

[-]

r-amp@reddit

In due time. Unfortunateky, western AI don't focus on open source. Too bad they don't get China's game.

Reply

[-]

Dreamthemers@reddit

I hope so. Probably if competition like Deepseek drops more open source models, it’s even more likely.

Reply

[-]

DAlmighty@reddit

I think Deepseek had its time in the sun but my bet is on Qwen being the real driver of open weight models.

Reply

[-]

createthiscom@reddit

I strongly disagree. If you care strictly about capability, Deepseek is still firmly in the lead. If you only care about what you can shove onto a single consumer GPU, Qwen becomes relevant.

Reply

[-]

DAlmighty@reddit

I respect that perspective, but I’m going off of the very short history that we have to go by. Qwen has been pretty consistently a top performer. Not saying Deepseek isn’t one, but Qwen has been one for longer in my eyes.

Reply

[-]

createthiscom@reddit

For coding capability, you only have to look at the Aider Polyglot Leaderboard: [https://aider.chat/docs/leaderboards/](https://aider.chat/docs/leaderboards/) DeepSeek is clearly the top of the pack for open weight models. I don't think people who use these models to code care about time in market or brand loyalty. I think they only care about capability.

Reply

[-]

DAlmighty@reddit

I feel Ike you’re missing the point, but that’s ok.

Reply

[-]

ReasonablePossum_@reddit

Calm down that ego dude. The guy told you their opinion and argued with proof, you went into adhominems, which means you not only lost the debate, but also exposed yourself as a fool.

Reply

[-]

DAlmighty@reddit

Bots gonna bot I guess.

Reply

[-]

ReasonablePossum_@reddit

Yeah... seems they just double down in their shame, and can{t learn.

Reply

[-]

createthiscom@reddit

I mean... is the point gooning? Because otherwise, yeah, I don't get it.

Reply

[-]

DAlmighty@reddit

Hahahaha I love here your heads at! No diddy

Reply

[-]

Xp_12@reddit

why did I read this

Reply

[-]

Gipetto@reddit

I need a version of that chart cross referenced with the amount of vram available so that I can see what the best thing that I can stuff in to what ram I do have.

Reply

[-]

HashPandaNL@reddit

That doesn't list the newest qwen models tho

Reply

[-]

UncleRedz@reddit

I think they are contributing in different ways, DeepSeek is mostly focused on research and have contributed tremendously to open source LLMs through their research papers and releases, perhaps by chance or luck, which in turn have influenced everyone else. And while its not visible from an outside perspective, they have changed the AI landscape inside China completely. It's less about number of releases and more about the impact they are having. Check out this article for a different perspective, [https://open.substack.com/pub/hanstakeontech/p/deepseek-6-months-later?utm\_source=share&utm\_medium=android&r=3jjyvu](https://open.substack.com/pub/hanstakeontech/p/deepseek-6-months-later?utm_source=share&utm_medium=android&r=3jjyvu)

Reply

[-]

forgotmyolduserinfo@reddit

Glm too

Reply

[-]

Fresh_Finance9065@reddit

Deepseek got set back by being forced to use Huawei npus, but they just got lucky with the gamble they chose with the MoE design being outstanding for Deepseek V3/R1. They still have the best model natively trained at 8 bits, so it quantizes better than other models which is the condition most providers run them in.

Reply

[-]

rz2000@reddit

Qwen has potential with other types of models, but I think it has serious flaws with the chat interface. It too quickly veers into talking about nonnegotiables and deadlines, and other versions of being emphatic about the point it is trying to make without putting in the effort to determine whether its analysis has drawn accurate conclusions. The problem could be the inherent vulnerability of Alibaba itself standing out as too big of a political target, so it necessarily adopts ugly traits when incorporating alignment features. Moonshot, Zhipu, and DeepSeek seem like they are putting out much less flawed LLMs right now.

Reply

[-]

silenceimpaired@reddit

Yeah, it seems more likely Deepseek existed to allow stock shorting and actual interest in LLMs is probably nonexistent for the group.

Reply

[-]

ttkciar@reddit

Only if Altman thinks he can leverage it to schmooze more rounds of funding out of OpenAI's investors.

Reply

[-]

usernameplshere@reddit

I don't think they will ever do it again, if there's not a major pr or reputation issue for ClosedAI

Reply

[-]

Upper_Road_3906@reddit

if musk isn't killed he will harass sam into releasing another version, I could see elon releasing an open source grok that you could use locally with your testla AI compute unit haha

Reply

[-]

ForsookComparison@reddit

> aka the gpu in your car I don't think they'd use the more-powerful FSD hardware for it and while their infotainment is way faster than other manufacturers, I don't think it's running any models decently. Could be wrong - but I don't see in-car (offline) Grok soon.

Reply

[-]

ElementNumber6@reddit

Kind of weird seeing Elon referring to himself by last name

Reply

[-]

ForsookComparison@reddit

Petty as it is I think this is a game of chicken between Xai and OpenAI. If Grok3 and its variants are made open-weight, then we'll see a new gpt-oss.

Reply

[-]

a_beautiful_rhind@reddit

I didn't feel safe enough with the last one. I hope the sequel doubles down.

Reply

[-]

yami_no_ko@reddit

Given that GPT-OSS is mainly a PR effort to tackle its "closed AI" reputation, this likely won't happen until they decide to address this issue again. They'd certainly brag and bla about a new effort at least 3 years in advance, but I wouldn't get my hopes up.

Reply

[-]

markole@reddit

I don't think that the reputation was the cause for the relese of the model even. Pretty sure they did it to handle the lawsuit from Musk.

Reply

[-]

cornucopea@reddit

GPT OSS is mainly a practical MoE for average joe kinda deal. Grok is still on the traditional scaling path, bigger the better and completely ignoring the edge market but racing to the "AGI" illusion. Very different strategy. Will OpenAI follow another OSS, that'll depend on how innovative the open source market is. With all these hundreds billion paramter models rolling out every day, there is no sign of the compelling of OpenAI to entice the open source market again.

Reply

[-]

epicrob@reddit

I think it is too early to tell. But I am sure it's not going to be as often as the updates of Chinese models out there. Yet, I am cautiously optimistic that they will still update their OSS counterpart from time to time for PR purposes.

Reply

[-]

SillyLilBear@reddit

100%

Reply

[-]

pigeon57434@reddit

there will definitely be a 2nd one but expect a slower release cycle similar to how long it used to take between Llama models (you know before llama died but it was many months between releases) so maybe a new one in december? but could be next year before the 2nd one

Reply

[-]

InevitableWay6104@reddit

I hope so, but definitely no where near as frequently as qwen. I think we’d be lucky if it’s once a year, if ever.

Reply

[-]

someguy@reddit

I think they only released it to stop getting dunked on for the "Open"AI name. I don't expect any more models from them in the near future, since they now have an irrelevant open model to refer to.

Reply

[-]

AnonymousCrayonEater@reddit

When it falls far behind they’ll probably release something fresh to stay in the top 10 but otherwise I think it’s mostly just to say they have open models.

Reply

[-]

thekalki@reddit

Looking at their repo for Harmony template which is invested with bugs and they are not even merging the pr from the community nor being maintained at all. So chances are slim anytime soon.

Reply

[-]

ConfidentTrifle7247@reddit

I suspect we will see one, but not with the cadence that the open weight Chinese models are released. Probably more of a every year or two release cycle. Mainly for optics.

Reply

[-]

NearbyBig3383@reddit

Oh brother, for the love of God, he just threw the crumbs to the poor that they threw because of that lawsuit that Elon Musk put behind his back, you know, something else, everyone was opening a linen model, they were the only air company that wasn't going to do it, you know, so basically that's it, that meal is a marketing ploy with a colleague on top, he said.

Reply

[-]

Unlucky-Message8866@reddit

Good model? Currently sitting at #77 at the gpu-poor benchmark HF launched this recent days

Reply

[-]

createthiscom@reddit

If they do, my only request is that they remove the fine tuning for their V4A diff format or whatever. It limits the model’s usefulness in large code bases.

Reply

[-]

AppearanceHeavy6724@reddit

OSS is a big success for local coding. They might repeat it next year.

Reply

[-]

No_Conversation9561@reddit

Anthropic is the last company I’d expect to be releasing open models.

Reply

[-]

AppealThink1733@reddit

Honestly, I'm not worried about that. I want to use the Qwen3 vl 4b, that's the one I want to test!

Reply

[-]

NoFudge4700@reddit

I hope GPT-4o one day. Maybe after 5 years lol. Or 10.

Reply

[-]

x0wl@reddit

GPT-OSS 120B is already at 4o level in text tasks, so you can just use that

Reply

[-]

balerion20@reddit

Sorry but it is not even same ball park, it is not even multilingual, you have to rl train afterwards.

Reply

[-]

SillyLilBear@reddit

Doubt it, since it was just a publicity stunt

Reply

[-]

Secure_Reflection409@reddit

It's basically king of 120b so I doubt we'll see anything for a long time.

Reply

[-]

shockwaverc13@reddit

no, unless we call them ClosedAI for 2 more years, only then they'll release something

Reply

[-]

silenceimpaired@reddit

I think it unlikely for some time. A year at least. I suspect they did it to decrease free tier usage, shut down those who said OpenAI isn’t open, and/or to demonstrate they have a stronger understanding of building local models than other AI companies so that potential clients will come to them for a tailored local solution.

Reply

Reply to Post

57 Comments