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

Models are back
by Strube, Alexandre 02 Oct '25

02 Oct '25

There were some changes on the authentication processes of UNICORE/Unity, and this kept the models which run on our supercomputers (Qwen3-235, GPT-OSS, Llama3.1 405b) out in the cold. This is fixed, and we’re back in business, baby! There are some special people at JSC. One of them is Bernd Schuller, the maintainer of Unicore. If you are reading this, thanks a lot! Keep barking, Alex

1 0

New models and api server changes
by Strube, Alexandre 02 Sep '25

02 Sep '25

Hello Blabladores! It’s been a while since I wrote last. Things haven’t been quiet, though. Many of you noticed that I added some models - like GPT-OSS, for example - and they do not appear on the API listing. That is not an accident. We have been preparing for some months a new, faster API server, which better capacities to support the upcoming new hardware we will get for Blablador. The models running on our cloud instead of the old cluster are pointing to this new server. At some point, all models will run into the new server, therefore the new server replaces the endpoint of the old one. If you want to check the models running on it, give it a try: http://134.94.199.27:8080/v1 This week and the next are pretty busy, so I most probably won’t change these right away. But this endpoint above will at some point disappear and we we’ll only have the default one, with all models. Another interesting thing is that we have the Hunyuan-MT-Chimera-7 translation model available. It’s the first time we have a model only for translation. I hope it’s useful! Let’s bark! Alex, el Blablador

1 0

New read_user scope for api token
by Strube, Alexandre 08 Aug '25

08 Aug '25

Hello Blabladores! Some of you mentioned that using the “Api” scope on helmholtz codebase for BLABLADOR was too much. Helmholtz kindly added the “read_user” token, so one can create a much more strict token and use it instead of the old “api” one. So if you want, just go to https://codebase.helmholtz.cloud/-/user_settings/personal_access_tokens and create a new token with this much more limited scope, and revoke the old one, and things should work the same, but without any access to anything on your Helmholtz Codebase! Let’s bark throughout the weekend! Alex

1 0

GPT-OSS is out
by Strube, Alexandre 06 Aug '25

06 Aug '25

If you heard the news, OpenAI finally released an open model. They say it’s good. I don’t know. You tell me. Because it’s on Blablador already! Less than 24h after launch. It’s not on the main api server yet, though. Use the experimental one: http://134.94.199.27:8080/v1 This one gave some trouble. OpenAi made a TERRIBLE job with the deployment of this one. They should learn with DeepSeek and Qwen, really. Give it a try! Let’s bark! Alex

1 0

AI and Research Data Management: What is your use case?
by Strube, Alexandre 31 Jul '25

31 Jul '25

Hello Blabladores! The Helmholtz Open Science Office and Helmholtz Metadata Collaboration (HMC) are currently planning the the second Research Data Commons event on September 30th, the topic being artificial intelligence and its application in research data management (e.g., practical perspectives on how AI tools can support FDM and/or how to manage data best for use in/with AI tools). It would be very nice if a scientist using Blablador or other AI tools for research and data management could give a 15 minutes talk there. As I don’t spy on you, I don’t know what you are doing with it :-) If you are interested in show what you do, it would be super nice! Please let me know! Alex

2 1

Alias-large got better!
by Strube, Alexandre 31 Jul '25

31 Jul '25

Alias-large was updated with the new Qwen3-30B-A3B-Instruct-2507, a non-thinking model, which is a significant upgrade from the April version. more news here: https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507 Qwen/Qwen3-30B-A3B-Instruct-2507 · Hugging Face huggingface.co Alex

1 0

Hello Kimi-M2, goodbye Kimi-K2
by Strube, Alexandre 22 Jul '25

22 Jul '25

As the competition for the best LLM doesn’t stop for a second, as soon as Kimi-K2 made splashes by being as good as some frontier private models, Qwen comes back with a revised model on their own: Behold Qwen3-235B-A22B-Instruct-2507! (those names are getting out of hand already, but it’s still better than whatever openAI calls their stuff) Both Qwen and Moonshot, the creator of Kimi, are backed by Alibaba, the e-commerce giant. The biggest advantage of Qwen is that it’s a fifth of the size of Kimi’s, which has a trillion parameters, while Qwen has 235 billion. This makes it way cheaper to deploy. Qwen has an Apache license, and Kimi, a modified MIT one. Another nice thing about the new qwen is that they now have separate Thinking and non- thinking models, which makes it much easier to use with agents, for example. I was trying to get Kimi to run on Blablador, but that would be ridiculously expensive, given that their minimum recommended hardware is 16x H200 with 141gb of vram each, which would be something like 6 JUPITER nodes, just for a single model. Makes little sense. So, I got the latest Qwen, released last night, running temporarily here! Happy barking!

1 0

Maintenance AND NEWS
by Strube, Alexandre 16 Jun '25

16 Jun '25

As you might’ve noticed, Llama-3 405 is out, and so is Qwen-Long when running on our Helmholtz-WestAI hardware. This is because the JURECA system is scheduled for maintenance tomorrow, and it’s not accepting long time jobs now. Sorry for the inconvenience. There are good news, though: there are two models right at the top of the list on the web ui: Phi and Qwen-Long, just running on the newer hardware. Phi is, at least for today, the default one on the list, so most people using the web ui will be using it, and will notice that this thing is FAST. I put it as the default to stress test the infrastructure. - Pros: We are testing the production instance of Blablador The Next Generation (R) - Cons: - We are testing, things come and go (the testing on the test instance is done - we’re going live, baby) - API does not show all the new models yet - Random glitches The infra is not complete yet, but soon one will be able to use it on the api, and use all the resources newer models allow: function calling, image support, etc. Hope you enjoy it! Let’s bark! Alex

1 0

QwenLong-L1: Towards Long-Context Large Reasoning Models with Reinforcement Learning
by Strube, Alexandre 05 Jun '25

05 Jun '25

From the huggingface repo: QwenLong-L1: Towards Long-Context Large Reasoning Models with Reinforcement Learning We release 🤗 QwenLong-L1-32B, which is the first long-context LRM trained with reinforcement learning for long-context reasoning. Experiments on seven long-context DocQA benchmarks demonstrate that QwenLong-L1-32B outperforms flagship LRMs like OpenAI-o3-mini and Qwen3-235B-A22B, achieving performance on par with Claude-3.7-Sonnet-Thinking, demonstrating leading performance among state-of-the-art LRMs. We have it on Blablador. I don’t have yet a perfect system prompt for it, so you might see some funny tokens there, like <|im_end|> and so on. In any case, it would be nice to see if it’s useful for you, and then I would keep it running. Please let me know! Let’s bark! Alex

1 0

Hi
by Na, Sehyang 27 May '25

27 May '25

1 0