July 2025 - Blablador-news - lists.fz-juelich.de

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