September 2026 - Blablador-news - lists.fz-juelich.de

Laya, an open - source JEV
by Strube, Alexandre 12 Oct '26

12 Oct '26

Jev[1] took the AI world by storm lately, even though it’s not necessarily the WOW people are making out of it. It can be extremely useful, though. In any case, 2 days after Jev came out, someone had an open-source equivalent of it: Laya [2]. And now, you can try laya too. It’s alias-laya on the api, the endpoint being identical to Jev. Here is an example: curl https://api.blablador.fz-juelich.de/v1/systemone \ -H "Authorization: Bearer YOUR_TOKEN" \ -H "Content-Type: application/json" \ -d '{ "model": "alias-laya", "state": "Write an extremely ultra advanced script optimizing multi-threaded CUDA kernels for matrix multiplication.", "questions": { "route_selection": { "type": "choice", "instructions": "Which capability tier is required to successfully fulfill this request?", "criteria": { "fast_tier": "Simple coding scripts, basic factual lookups, and clear formatting tasks.", "heavy_tier": "Highly complex logic, multi-threaded optimization, math proofs, programming, or ambiguous system architecture queries." } } } }' | jq And the reply will be something like "model": "laya-rl-agent", "answers": { "route_selection": { "type": "choice", "choice": "heavy_tier", "probabilities": { "fast_tier": 0.4083, "heavy_tier": 0.5917 }, "confidence": 0.0244, "action": { "act_probability": 1.0 } } }, "usage": { "input_tokens": 82, "output_tokens": 0 }, "routing": { "model": "english", "repo": "convaiinnovations/laya", "reason": "English Latin text", "detection": { "script": "latin", "script_profile": { "latin": 1.0 }, "language": "en", "is_english": true, "language_undecided": false, "diacritic_rate": 0.0, "non_latin_fraction": 0.0 }, "workflow": null } } And this query will run in… I don’t know, 0.05 seconds or so. Hope it’s useful! Have fun. [1] https://en.wikipedia.org/wiki/Jev_(AI_model) [2] https://huggingface.co/convaiinnovations/laya Let’s bark! Alex, Blablador honcho

2 2

Juwels booster's jobs are unreachable
by Strube, Alexandre 30 Sep '26

30 Sep '26

While they are all running, unfortunately they can’t be reached, so all the big models are not reachable. Should be back very soon

1 0

Helmholtz codebase is offline - we have no tokens
by Strube, Alexandre 28 Sep '26

28 Sep '26

Yes, things are offline now at the Helmholtz Codebase at DESY, and we are on it. Sorry for that Alex

1 2

Laya (Jev)
by Strube, Alexandre 25 Sep '26

25 Sep '26

Laya has moved to a better location. https://laya.blablador.fz-juelich.de <https://laya.blablador.fz-juelich.de/> Please report any issues on our Matrix channel on Blablador! Mind you, it’s experimental and we are messing with things. Don’t expect 24/7 availability just yet. Alex

1 0

Booster down - many models down
by Strube, Alexandre 25 Sep '26

25 Sep '26

There’s a huge update on booster, and the models are being restarted. I was just informed. Sorry

1 0

Router failure: laya will have to wait
by Strube, Alexandre 24 Sep '26

24 Sep '26

While I have laya running, our current router doesn’t like it, and we will come with a proper solution. Sorry to have hyped you up. I should have know better. By the way, laya is REALLY light. You can try on your laptop without a gpu: uv pip install laya Sorry for the blunder, Alex, the one which barks and bites himself

1 0

Blablador on JSC supercomputers
by Strube, Alexandre 24 Sep '26

24 Sep '26

Just a heads up: If your project needs some AI input, Blablador is accessible from the compute nodes, the same way it is from your laptop. The same endpoint, the same api key. Just give it a try! Best, Alex

1 0

Experiments: New models and modalities
by Strube, Alexandre 22 Sep '26

22 Sep '26

Hello Blabladores! If you have been using the system lately, new models keep popping in and out of existence. Mostly in. So, what’s new? - DeepSeek came out with a new v4.1 version. It’s so good it should be called 5.0. It’s there, it’s good and runs fast. - Qwen3.8-next flash is amazing. So good that it’s our recommended alias-large. - Kimi-K3 is there for a while, but it’s finally usable. If I would be an economist, I would say I got a 22400% improvement on it. 8x on speed, 2x on context size, and 16 times on concurrency, with roughly the same hardware. - New Qwen 4b embeddings and rerankings: they are there for a project, but can be used by anyone, and they are fast. Same for 8b versions. - MiMo 2.6: It was released this morning, on Flash and Pro releases. It’s the best open source model out there. And it’s running since today on Blablador, on both versions. It’s amazing, actually. It also is multi-modal. Now, for the quirky ones: Anonymizer: M-size ShinrAI PII 1.3: This is pretty cool. You feed a document with personal data, and it anonymizes it. Example: If I feed the model with "I, Cornelius Gildemeister, from the Değirmen Yolu 69, have the credit card 4322 3413 3321 3434 and live in Neuwarden, Germany" The model will do this: I, Alperen Tunçbilekli, from the Fıstıklı Bahçe Sokak 40, have the credit card 2221 8355 9496 6697 and live in Obermoosbach, Germany — 3 entities (115 ms) — • PERSON: İlker Ovacıklı → Alperen Tunçbilekli (model, 0.63) • STREET: Kuyumcular Çarşısı 123 → Fıstıklı Bahçe Sokak 40 (model, 0.76) • CITY: Kaltenhagen → Obermoosbach (model, 1.00) This is very nice to anonymize documents, and it’s fast. Image generation The models are running there for a while. We have at the moment 3 versions: FLUX2-Dev, and two Qwen models (we are still looking at licenses). There is tuning to be done at them, but I’m happy it works and since now, it works on the main website. The Qwen-image-2.1 is also an image editor, so you can upload an image and ask it to change it. You can also set the resolution and number of steps on the prompt. So, something like "A potato, 256x256, 10 steps” will be much faster (and worse, and smaller) than simply “a potato”. A little screenshot: No, I won’t remove the watermark. It’s the law. We deal with it. I tried an example of image editing on it: Sent a picture of me and said: “make this person into a Klingon”. Somehow I became a Greek singer, I guess. (hint: qwen-2.1 image accepts multiple images as input). Well, I think this is it. Just some status update. This is all experimental, but it’s ready to use. Hope you enjoy it. Let’s bark!

1 0