OpenAI releasing new open model in coming months, seeks community feedback
81 comments
·March 31, 2025vessenes
verdverm
How about everything needed for full reproducibility?
This is what I would like to see
vessenes
Seems unlikely. :) As is an end to end multimodal model, I'd guess. But we can ask!
nickthegreek
Id hope they will also start to target machines with 128gb of unified ram now that we seem to have at least 3 options on that front.
politelemon
I'd rather see more openness and the ability to run on commodity hardware. There are hundreds of options on that front.
impossiblefork
I think that when learning privacy and being able to run things oneself is less of a problem.
Keeping ones work confidential becomes interesting once one starts doing new things, so while it would be fun, I don't think that's the most useful type of open model. I think code models are probably the most useful.
EGreg
and guess what
they did announce they'll release open weights :)
https://x.com/sama/status/1906793591944646898?s=46&t=6NqVriD...
to be honest many of us never saw that coming... LOL
vessenes
I don’t think it will be an end to end multimodal model, unless they’re holding that shocker for later - he says “language model” in the announcement. So basically o3 mini, not 4o multimodal.
syntaxing
Easy. A reasoning model with better performance than QWQ but at 21B (like Reka Flash 3) and good tooling call support. A model as “intelligent “ as Qwen2.5 but personality and creativity of Gemini (or Gemma at a minimum)
rcpt
And also you should get prizes for using it
xiphias2
Something even cooler would be a model trained for 4 or less (1.33) bit weights instead of quantized after pretraining.
Math units are completely underutilized when I'm inferencing with batch size of 1, and post-training quantization under 8 bits loses too much of the precision to make a real difference compared to smaller models with higher precision.
pera
Open model is such a misnomer: it's like calling an ELF an "open executable".
Distributing things for free doesn't make them "open". The reality is that free (as in free beer) weights are closer to freeware than anything open source. In fact, since all these models are build using pirated media a more appropriate term could be plain old warez.
WithinReason
Getting weights without the training set and training scripts still gives you a form that's modifiable by end users, a single person can fine-tune a model. Getting the training scripts and dataset gives you nothing useful unless you have millions to burn. "Open weights" are closer to the spirit of open source than training scripts and datasets.
kgeist
Current open-weights models are not as multilingual as GPT3 or GPT4. I'd like to see support for more languages.
bigdict
Gemma 3 is.
bhouston
Q: How open?
Can it be used commercially? Is the training protocol going to be open? Or just the weights released like Llama models?
minimaxir
> Can it be used commercially?
OpenAI has been good about sticking to the commercially-friendly MIT License for their OSS. (e.g. Whisper/tiktoken)
fooker
It's a kind of an open secret that there's no 'training' protocol for these state of the art models.
Researchers behave like alchemists when training these models, and the actions are not really reproducible.
rockinghigh
They could provide access to the training code. It's useful for training smaller models or distilling larger ones. They don't need to release every details involved in tuning the optimization parameters during the pre-training stage.
fooker
There is no training 'code' that will get you anything close to a usable result.
nightski
At this point I think it is safe to assume open means open weights only.
n2d4
They specify "open-weights" later down the form
simonw
That doesn't mean anything in particular other than that the weights can be downloaded and run. What's interesting is the terms under which they can be used - there are quite a few "open weights" models that can only be used for non-commercial purposes without a paid license, like Command R from Cohere (cc-by-nc-4.0).
mbrezu
> Let's start with your details
paxys
In this case I don't see a problem. If you want your opinion to hold weight then put your name/background/credentials/reputation behind it. Not every discussion needs to be fully anonymous and upvote-based.
vntok
Well obviously they will have thousands of applications from which they need to make a selection removing trolls, luddites, etc.
How would you do it without asking for some applicants' details?
chad1n
Considering how much they trust their LLMs, why don't they just run o1-pro to make a summary of the responses given in the feedback
lawlessone
>they need to make a selection removing trolls, luddites, etc
Do the luddites opinions not count?
wewtyflakes
If you have a business selling hamburgers, does it make sense to ask vegetarians what should be on it?
vessenes
Well traditionally you're supposed to make them count if you're on the luddite side.
stale2002
If their plan is to sabotage things, then no their opinions should not count.
vntok
To the tech sector's asking for input to help shape their vision? I would guess they don't.
mirzap
OpenAI should release its frontier model as an open-weight model. There are already open-weight models that match OpenAI's best models (at the time of their release), so the idea that OpenAI would lose something by making its frontier models open-weight doesn't hold up. With an open-weight model, they would instantly kill any proprietary competitor, similar to what Google did to all its competitors with Android.
IMO, OpenAI should focus on tooling, infra, and setting standards for AI apps and profit from those. MCP is what "custom GPT" is supposed to be; OpenAI lost that battle, among many others.
Gopher started as a freely available protocol but later changed its licensing terms, requiring fees to be paid—similar to what OpenAI has done. We know how Gopher ended up: today, most people haven't even heard of it, despite its adoption at the beginning.
dilippkumar
> There are already open-weight models that match OpenAI's best models (at the time of their release)
Did I miss something? Do we now have o1-pro level performance in an open source model?
nyrikki
Depends on your use case.
DeepSeek R1 is far better at producing maintainable, modularized code as a coding assistant as an example.
The big deal is that the distilled versions like DeepSeek-R1-Distill-Qwen-32B are good enough that anyone with a few old 1080 Ti's sitting around can run them and get most of the performance.
When you can run gemma3/qwq/DeepSeek-R1-Distill-Qwen-.../etc... you can easily switch models when one fails too.
And you have consistent performance that doesn't degrade over time, have the ability to avoid leaking prompt data between client etc...
It is all horses for courses though. For me o1-pro is roughly the same as o1 with just higher limits etc... but is still worse than o1-preview IMHO.
In my experience the few percentage points on synthetic benchmarks that o1-pro was claimed to have doesn't matter much in real world problems.
R1 pretty much matched o1-1217 on every benchmark and the distilled models like DeepSeek-R1-Distill-Qwen-32B only lost a tiny fraction.
A few months of o1-pro costs will get you a local usable model of GPUs if you are fine with ~20 eval tokens/sec.
But if o1-preview wasn't better for your use case than o1-proe...the calculus can change.
maxloh
Not yet! But we have some close ones. For example, OLMo 2[1], OlympicCoder[2], and OpenThinker[3].
[1]: https://allenai.org/blog/olmo2-32B
[2]: https://huggingface.co/open-r1/OlympicCoder-32B#evaluation
bilbo0s
There are already open-weight models that match OpenAI's best models (at the time of their release), so the idea that OpenAI would lose something by making its frontier models open-weight doesn't hold up
If true, that cuts both ways though right?
It also means we lose nothing if OpenAI releases a closed weight model.
Liwink
What's the major use case of open source models? The stable diffusion community seems pretty active. A lot of fine tuning to generate NSFW. What about LLM?
minimaxir
The biggest use case of open-source LLM is that your use of it is private and input is not sent to third-parties. For a personal perspective, it means it's free and accessible forever (and yes can be used for NSFW stuff), and from a business perspective, it matters both for legal reasons such as finetuning on propritary data and it mitigates the business liability of issues with the third-party LLM provider such as random API outages.
daemonologist
It's a huge compliance hurdle at my company to add any new vendor, so we almost exclusively use open models that we can run on our own hardware (or our rented cloud instances). Even just getting Bedrock enabled in one of our existing AWS accounts has been in the works for months.
ewalk153
When coupled with smaller parameters sizes, it enables BYOD, different cost scaling, and local inference.
turnsout
This definitely feels like a move made out of desperation, but I admit I'm curious to see what they release.
It's fascinating that some of the best-funded startups of the 2020s are all rushing to commoditize their core technology as quickly as possible. They all seem to have the same business model as the Change Bank from SNL
minimaxir
Desperation? How so?
The cynical late-stage-capitalism argument for releasing an open source model is for PR/goodwill and hoping for the chance the model becomes a foundation model in the OSS community. But OpenAI definitely isn't operating around good PR and the chance of releasing a foundation model is unlikely given the competition.
toomuchtodo
Orgs and individuals are chasing sentiment to cash out while valuations are inflated (there has been some expectation deflation, see Microsoft terming datacenter leases and the CoreWeave IPO fizzle), but open models are still welcome (democratization).
turnsout
Desperation because they face an existential risk from trillion dollar companies (and others) creating open models that are starting to catch up to their flagship.
null
CamperBob2
What would you like to see in an open-weight model from OpenAI? Explain what you would use it for
See... that's kinda the idea behind an open model. I don't have to explain what I would use it for.
jug
Sounds like an open GPT-4o / o1 distill or something… And that they want to know what to go for.
bigyabai
The idea that they want to train a new custom model for open release instead of just... giving us GPT-3 already suggests a terrible start. I'm calling it now, this is a strategic counterplay against Google's Gemma models so @sama can sell Tim Cook a "frontier" local model that doesn't compete with anything coherent. A fig leaf for their "Open" identity and a paper tiger for the Apple Intelligence panoply.
OpenAI doesn't believe in Open Source, they merely want it's prestige without committing to it on-principle.
minimaxir
> instead of just... giving us GPT-3
If you're referring to the GPT-3 from 2020, modern open source models five years later are a) better at benchmarks b) much smaller yet still better at said benchmarks c) much, much cheaper/faster due to architectural improvements.
The real hard thing for OpenAI to do is to release an open-weights model that's better/more differentiated than Gemma 3 (at the small scale) or DeepSeek R1 (at the large scale)
mackenly
But if OpenAI was Open, they'd open-source those old obsolete models. You're right that no one really wants it, and it has little to no commercial value at this point, but that's all the more reason they should just put it out there and actually live up to their name.
null
refulgentis
Is there anything about training a new model? I just assumed they were asking for all the little bits that companies forget to do at release with an open model.
null
anxoo
[flagged]
hashemian
The original comment says "their mentality about open-model is wrong, evident from their post". Does not say it's shady! Why to be cynical?
koakuma-chan
I'm surprised they expect people to voluntarily undergo that interrogation, for their feedback to then be processed by some LLM.
vntok
If you're interested in the field, why wouldn't you answer the questionnaire? It costs you basically nothing while the upside is getting something that is potentially at least a tiny bit more useful to you than if you hadn't said anything?
42lux
Don’t kid yourself about the intentions. It’s only so enterprise customers can deploy a shitty chatbot on premise for their "secret data". The models will be free for commercial use until 1 mil turnover or something like that.
Spivak
I mean for large enough buyers that was already an option, they wouldn't need to release a new model just for that.
rockinghigh
I don't believe OpenAI provides on-premise serving.
42lux
Citation needed.
For those wondering how to answer "what do you want to see from an open model" I put this in: an open weights end to end multimodal model, a model large enough to act as a functional teacher along with a range of nicely distilled smaller sizes, code repo to make training / finetuning easy.
As I write this, I'd also like to request a set of tool calling LLMs in various sizes. Feels to me like a small fast local tool calling assessor with a large context to support a lot of MCP functions would be very useful locally.