Open-source or not?
I'm really confused with the licensing model. In the arxiv paper, https://arxiv.org/abs/2607.09424 it's titled "A Sovereign, Open-Source Foundation Model for German and English" but the license in Huggingface is clearly non open-source. Can you clarify this? Thanks a lot.
I've never seen anything quite like that.
The entity behind the "German research consortium" made claims about the openness of the model before the public release, plus the model already sits right there in "public" access.
But nobody can download the model because of the "gate", its card is full of placeholders (e.g. TODO: confirm maximum context), and the license's set to "other".
Just tested the base model (Nemotron-3-a3b-30b) and since im gated i post my response, its just stupid, a3b cleverness is comparable/less to gemma-4-e2b a micro 5b model.
In my opinion if the experts are too small you can train like hell, it keeps 3b and acts like 3b.
Would say the baddest expierience since im into llm. Guessing death by birth...
No chance to get bet access. We mailed, contacted the people and their linkedin, but no reaction. Then we complained publicly in linkedin - no reaction.
We are a German entity and do the https://ypipe.com AI orchestrator/ coordinator. Our focus are normal smaller models and to have them working in flows and such with MCP servers.
Clearly, with Soofi we have the chance to run a German model in a German orchestrator - so it matters for us to test it and to adjust our software that in runs in it as soon as possible (even if it is larger than our normal scope) - but even that does not help.
It is funded by 20 million so far by EU and as far as I got it made by Telekom.
Goal: Take the public money, block everyone out.
Now there are benchmarks celebrated and nobody has access to the model.
I am really really sad about that - we were excited in the first place, but now it is agian this non transparent and hidden club thing.
To those tracking this issue, Soofi updated the README file 4 hours ago,
⚠️ Closed Beta Release Disclaimer
This model is a beta preview and a research artifact. It is not an open release.
This checkpoint is an intermediate result, currently in a closed beta with already selected partners to work on an improved version. It is under development and subject to change based on partner feedback. Outputs, weights, and behavior may differ between checkpoints.
The final model will be released openly under a permissive license, without gated access. We will share access details as soon as it is ready.
TL;DR the model is still in development and it's unknown when it's going to be available to the public.
comparing to the old one , that I saved it is not a lot different. didn't they say everybody who request access will become 100% access? confirm frisky?
comparing to the old one , that I saved it is not a lot different. didn't they say everybody who request access will become 100% access? confirm frisky?
No clue, sorry! Since it's a gated repository, we can't check the previous commits, only the latest one.
i saved last README -> true. can u find researcher message about access? i cant find anymore ... frisky?
It seems... just google for Airport BER... Well planned, well financed, absolute fiasco. As always if a Federal Ministry is wasting funded millions.
TBH, for some Local AI, maybe the casual gamer this model and size is okay.
But really NEMOTRON-3-a3b-30b for the INDUSTRY? I'd rather choose Gemma-4-a4b-it for my smartphone cause it gets what i want, a3b not.
REALLY? A simple google:
"Nemotron model kinda sucks."
"Out of the box nah."
"Weak and prone to hallucinations."
"Sehr meh."
A simple i do this then this after that this, needs at least 3-5 regenerations to get it understand what i want. and thats exactly where i place the a3b. between Gemma-4-e2b and e4b but the gamma's are 5b/8b and their understanding is far better they get straigt to the point always.
WHO HAS MADE THE MOST STUPID DECISION I'VE EVER HEARD OF. Maybe: you get free A6000's decision!?
My bet, the selected Industial partners are already faceplaming so they fear the community to destroy them completely.
We hear you, and we truly appreciate the incredible interest in Soofi S.
The number of requests for early access, testing, and technical details has exceeded our expectations. We know many of you are eager to get your hands on the model.
The models on HuggingFace are intermediate results and in a closed beta with already selected partners to work on an improved version. Soofi S is currently in the final post-training stage. We are working hard to release a model that is not only open, but also robust, reliable, and genuinely useful for European businesses. Our final checkpoints will be freely available (with a permissive license), without any gated access.
Our target is an open release within the next weeks and will share access details as soon as the model is ready.
Thank you so much for your enthusiasm, your patience, and your support.
I understand you have a lot of work right now, but directly answering the community's questions would go a long way in helping everyone understand what is happening behind the scenes.
Let us focus on the specific questions that have come up so far in this thread:
Why does a project heavily funded by federal/EU taxes require extensive personal information just to request access to a public foundation model?
Architecture Clarificatios were asked.. are they resolved by the benchmark paper?
What are the actual approval criteria for the closed beta? Oher gated repositories (like Meta's) usually grant automated access immediately?
Would be awesome to get some information here and not a generic political answer that just tries just to postpone..
Thank you in advance for all efforts...