NXT EU (@NXT4EU)
Germany has launched one of the world's best open-source AI models. Soofi S, made by the Soofi consortium, is a 30B parameter model fully trained in Europe and tops the ranking for open-source AI. Huge moment for Europe, and finally some competition for Chinese open-source AI.
Jul 15, 2026 · 10:44 AM UTC
214
583
4,776
423,946
Sort replies: Relevant Recent Liked
[](https://xcancel.com/effi288/status/2077382659743924690#m)
Replying to @NXT4EU
Thanks for the coverage! 🙏 Two quick corrections from the team: (1) this is an outdated version of the figure, the Qwen3 32B point was an error we've since fixed, current version below (2) To be precise: Soofi S is the strongest fully open model (weights, data accounting, code all released), Qwen3.5 is ahead on raw capability; our edge is capability + throughput + full openness
30
22
396
18,219
[](https://xcancel.com/NXT4EU/status/2077383397425459233#m)
For a first model, very impressive. Excited to see what the future of Soofi will be.
4
109
14,268
[](https://xcancel.com/vr8vr8/status/2077468315992100969#m)
Replying to @NXT4EU
List missing all Qwen 3.6 and Gemma 4. And many other fine tuned Qwen 3.5. So its compared to things that already irrelevant ...
63
[](https://xcancel.com/realalexniebuhr/status/2077394461907173717#m)
Replying to @NXT4EU
@0xSero @MiaAI\_lab any chance you can benchmark the model?
1,439
[](https://xcancel.com/HorveiOYVIND/status/2077463359465640114#m)
Replying to @NXT4EU
Another proof that EU is not the "sinking ship" that anti EU far-rights-wing in Norway claim.
1
1
18
[](https://xcancel.com/PsMcGrath/status/2077399489111798210#m)
Replying to @NXT4EU
Congrats to the Soofi team, excited to see what it does on LMSYS Arena or Hugging Face Open LLM Leaderboard
185
[](https://xcancel.com/MightyBuddha01/status/2077398412266266671#m)
Replying to @NXT4EU
But can it run Doom? /s
1
2
415
[](https://xcancel.com/Saurax20/status/2077458191248437650#m)
Replying to @NXT4EU
Open source, does it mean it's free to download?
1
1
1,238
[](https://xcancel.com/ndeville/status/2077397286905057301#m)
Replying to @NXT4EU
@grok can this run on a Mac Studio?
1
5,336
[](https://xcancel.com/Zanmato%5F13/status/2077426272196514238#m)
Replying to @NXT4EU
Even as a French i'm happy to see some German models finally. The Germans were very good at Music so this can carry over to AI 😄
3
1,858
[](https://xcancel.com/rotorB/status/2077428774291816459#m)
Replying to @NXT4EU
There is no comparison with Qwen3.6 on the chart.
1,401
[](https://xcancel.com/A%5FA%5FToulkeridis/status/2077397035922190443#m)
Replying to @NXT4EU
i can't take seriously a plot that shows Mistral as Ministral three times
354
[](https://xcancel.com/bitoshimoto/status/2077410880929579093#m)
Replying to @NXT4EU
r u kidding me? i can't even test it. Looks like a scam website with lots of EU Logo. Prove it or it doesn't exist. Otherwise you are just another call center scammer
67
[](https://xcancel.com/Claude3Damme/status/2077363902518587405#m)
Replying to @NXT4EU
It is good but you have to obtain Großsprachmodellnutzungsgenehmigungsbehördenerlaubnisbescheid from the government to use it.
6
4
222
10,144
[](https://xcancel.com/FuAnTaJi/status/2077426716956319811#m)
Replying to @NXT4EU
All we need is Qwen3.7. Preferably, 122B-A10B.
202
[](https://xcancel.com/Einkaufskorbens/status/2077408788471017607#m)
Replying to @NXT4EU
Many stupid comments. Actually a really decent model and especially interesting for german industrial applications.
73
[](https://xcancel.com/fabianschierz/status/2077371901035384983#m)
Replying to @NXT4EU
You‘re comparing it to open-source models that are already outdated…
1
74
4,178
[](https://xcancel.com/donel%5Ftramp/status/2077415883446653303#m)
Replying to @NXT4EU
red flags everywhere… again some real world underperforming crap that has been funded by tax payers money 🤦🏻♂️
2
934
[](https://xcancel.com/DineshKumarOFC/status/2077449344098467933#m)
Replying to @NXT4EU
i thought Mistral is the only european ai lab, but theres few more
956
[](https://xcancel.com/xalturicx/status/2077395652850364708#m)
Replying to @NXT4EU
I'm purely curious, and I would ask this about any other lab, why do you benchmark against older models? i.e. Qwen 3.5 and not 3.6 for example?
2
1,091
[](https://xcancel.com/ape2616/status/2077439175817621760#m)
Replying to @NXT4EU
Dont worry, they'll keep a cap on it
53
[](https://xcancel.com/MisakiAnimated/status/2077393893381869893#m)
Replying to @NXT4EU
The most important question for me, is its training data ethical? For me, performance is not an issue, as long as no copyright infringement has occured. That is the model I'll push and support without reservation.
1
1
1,296
[](https://xcancel.com/andrei%5Fno%5Fno/status/2077398726759297145#m)
Replying to @NXT4EU
These guys call themselves "sovereign", but I guess they don't understand the concept. EU/Germany does not produce independently any of the chips or equipment necessary to train such a model.
1
785
[](https://xcancel.com/hypertectonic/status/2077389034783920452#m)
Replying to @NXT4EU
i'm sorry, i really don't mean to be harsh, but hearing "European model" is just cringe.
1
23
1,724
[](https://xcancel.com/fritz1707/status/2077393906312905042#m)
Replying to @NXT4EU
its still a small, not as capable model but i love seing people in eu trying it!
277
[](https://xcancel.com/JorgeCa14980958/status/2077398867411140736#m)
Replying to @NXT4EU
@grok how does it compare to Claude and Open AI models?
1
2
5,284
[](https://xcancel.com/jareddavisoba/status/2077414606780154275#m)
Replying to @NXT4EU
Good to see Europe getting involved and creating healthy competition.
1
4
1,818
[](https://xcancel.com/Bassmaster187/status/2077410461721514031#m)
Replying to @NXT4EU
Well, that doesn't seem to me to be very "open-source".
3
3
44
2,330
[](https://xcancel.com/thyrrium/status/2077394654199226680#m)
It's a nice start but nothing more. This is in no way anywhere near the top except for on their own metric catered to their model. The tweet hypes this up way too much, it's silly. All the best to the team though! Let's hope they actually get there
4
448
[](https://xcancel.com/tiberius%5Ftalos/status/2077471233369932261#m)
Replying to @NXT4EU
I run Qwen 3.6 on my 5080, the graph compares it to even older 3.5. This is not competition...
10
[](https://xcancel.com/oscargaske/status/2077391759512883615#m)
Replying to @NXT4EU
Is this a competitor to GLM 5.2? I don't see it included
1
4
3,359
[](https://xcancel.com/TheReapsZ/status/2077365087602389260#m)
Replying to @NXT4EU
Does not even have qwen3.6 27b on the chart
1
1,420
[](https://xcancel.com/Howaboua/status/2077366567868760422#m)
Replying to @NXT4EU
Sure, what was the actual model it was based on? I see configuration\_nemotron\_h.py & modeling\_nemotron\_h.py in the HF repo. I did send an approval request to check other files to have an idea. I have doubts, happy to be proven wrong.
3
42
6,685
[](https://xcancel.com/ECentrist71532/status/2077354915089748391#m)
Replying to @NXT4EU
This graph doesn’t have Qwen 3.6 35B, that being said Soofi is a great model.
47
3,879
[](https://xcancel.com/g4m3r0ps/status/2077394511232151724#m)
Replying to @NXT4EU
So basically you picked up the models you needed to make that Soofi S garbage appear as the best one, then ignored every other one that is better than it. This propaganda account is quickly becoming a joke 😆
8
607
[](https://xcancel.com/Huiche2025/status/2077390845431713997#m)
Replying to @NXT4EU
Where is GLM and DeepSeek? This looks deliberately engineered
1
1
527
[](https://xcancel.com/NXT4EU/status/2077343621011935648#m)