- Blablador-news - lists.fz-juelich.de

Ghristmas Mrrey!
by Strube, Alexandre 26 Dec '25

26 Dec '25

I found this jacket in a clothes stall at one of the Vienna’s Christmas market couple years ago. I regret to this day not having bought it. Anyway, I’m writing to let you know that NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 is running since day one, and from tonight, Minimax 2.1 is replacing Minimax 2 as our alias-huge model. Have some rest and go hug someone. Let’s bark, Alex, el Blablador

1 0

And we are back!
by Strube, Alexandre 03 Dec '25

03 Dec '25

So, the updates on the JSC infrastructure went pretty smoothly, and we are already back for some hours! While some instability might happen tomorrow, for now, we’re barking as loud as ever! PS: Mistral released their latest models, Mistral 3, yesterday. It’s already on Blablador for you to try with, and with function calling! Keep barking! Alex, el Blablador

1 0

Big models unavailable Wednesday and Thursday
by Strube, Alexandre 01 Dec '25

01 Dec '25

Dec 3, 2025, 07:00 - Dec 4, 2025, 19:00 A planned maintenance for the network backbone at the Jülich Supercomputing Centre (JSC) datacenter is scheduled from 07:00 CET on 2025-12-03 to 19:00 CET on 2025-12-04. During this period, all HPC, data, and cloud systems at JSC will be unavailable. Sorry for that :-( Alex, el Blablador

1 1

Blablador is DOWN
by Strube, Alexandre 03 Nov '25

03 Nov '25

The whole VM is down. I can’t access it myself. I’m going to the institute as soon as I can. Sorry for that. Alex

1 0

Phasing out Llama3-405
by Strube, Alexandre 27 Oct '25

27 Oct '25

Hello Blabladores, in order to improve the reliability of BLABLADOR, I am having to do some drastic changes, which required me to shutdown pretty much every single model and the server itself. OF COURSE while this happened, some GPUs decided do fail, which forces a full node reboot. I am bringing everything back up. If you are not working at the evening, just ignore this part. They are basically all up. Another thing is that I would like to phase out Llama3-405. It’s an obsolete model, it’s slow, and takes too many resources. Yesterday, Minimax-M2 came out, and it’s one of the best models out there. It’s ranked among models such as ChatGPT 5 and Claude 4.5. It is already running. It’s aliased as “alias-huge”. But as this is a reasoning model, you might filter what comes between the <think> and </think> on your code. Another thing is that this model might be able to generate images and music, but I need to figure it out. Let’s bark! Alex

1 0

Blabla fork for private use?
by Mario.Muellhaeuser＠dlr.de 17 Oct '25

17 Oct '25

Dear Alexandre, thank you for providing Blablador! I use it frequently for business. Now I am looking for a privacy friendly AI for friendly private use. I searched the internet and only found quite costly providers like mindverse, offering 250 tokens per month for 19 EUR. Can you suggest a more affordable AI aas provider, which offers privacy? Or are there plans or possibilities to open Blablador or a clone for public use, of course for a fee? Regards Mario -------------------------- 20 Jahre Forschungshubschrauber ACT/FHS - DLR Portal<https://www.dlr.de/content/de/artikel/news/2022/04/20221202_20-jahre-forsch…> DLR-Forschungshubschrauber: 20 Jahre Innovative Hubschrauberforschung - YouTube<https://www.youtube.com/watch?v=z_7RL3IsGrs&t=211s>

1 0

And we are good!
by Strube, Alexandre 16 Oct '25

16 Oct '25

Seems that most of the problems had been solved for now! Uhuu!

1 1

Outage of systems
by Strube, Alexandre 16 Oct '25

16 Oct '25

Hello Blabladores, you might have noticed that some models have been changing, coming and going. As they say, troubles like to come with company :-) So we have a couple issues together: 1 - The supercomputer where we run 4 of our models (GPT-OSS, Qwen3-235, Llama3-405 and Qwen-3-Coder with function calling) is offline. You can check the status of Jureca-HWAI on https://status.jsc.fz-juelich.de/ 2 - The change on the api server last friday. Given that we use mostly the VLLM backend, and VLLM has changed architecture recently. As mentioned here, https://docs.vllm.ai/en/latest/configuration/conserving_memory.html#quantiz…, "CUDA graph capture takes up more memory in V1 than in V0.” This in turn made many models run out of memory on the same hardware they ran before. So I am carefully reducing context size and the size of cuda graphs. It’s a manual, boring and slow process. I am sorry, working as fast as I can here so we can keep barking loud!! :-D Dr. Alexandre Strube a.strube(a)fz-juelich.de Helmholtz AI Jülich Supercomputing Centre Forschungszentrum Juelich GmbH 52425 Jülich, Germany Phone: +49 2461 61-3866 JSC is the coordinator of the John von Neumann Institute for Computing (NIC) and member of the Gauss Centre for Supercomputing (GCS)

1 0

Change of API server
by Strube, Alexandre 09 Oct '25

09 Oct '25

Hello Blabladores, many of you have heard that I have been testing a new api server on the past few months. It’s pretty stable, very performant, and it allows us to do way more with the models we have been running. So far, it has been on a http instance internal to FZJ, but I plan to move all models to it tomorrow (Friday), so this will be our official endpoint. If you are using the api.helmholtz-blablador.fz-juelich.de <http://api.helmholtz-blablador.fz-juelich.de/>, you don’t need to do anything. Only that more models will appear there. If you have been using the ip address of the secondary api server, just revert to the normal one. All models should appear there. Yay! Let’s push into production on a Friday, after all, WHAT COULD POSSIBLY GO WRONG? Let’s bark! Alex

1 0

Function Calling!
by Strube, Alexandre 09 Oct '25

09 Oct '25

(1) (gemini can’t write in vertical). the other models are worse Hello Blabladores! I have been trying around with enable models for function calling, as I this is something I wanted to do for a while. Please have a look at the model on the second api server - the name is "Qwen3 Coder 30B with function call” Once I get it into a stable working, it will be the new alias-code. I would like to run Qwen3-480, but it’s just too big and expensive to run. Let’s bark! Alex (1) Original picture from https://en.wikipedia.org/wiki/London_Calling

1 0