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

As claude goes offline, Blablador usage peaks
by Strube, Alexandre 03 Mar '26

03 Mar '26

If you have been using Blablador the last two days, you might have noticed slowdowns and even timeouts in some models. The usage raised A LOT lately because of software agents. While the number of unique IPs is roughly similar to a couple months ago, usage per unique ip has raised dozens of times. Another issue is that all the latest models do the thinking before answering. This has been a PITA, given a simple “hello” might take dozens of seconds of the model thinking about its own life before answering. For those playing with OpenClaw and Opencode, ClaudeCode and so on, and feel upset with the response speed, we are doing our best on our side - we have been caching requests where possible and scrambling the queue and adding scale where we can - but in the end, we have limited hardware. If you are upset with this, there are free alternatives on OpenRouter. I checked and they have half a dozen alternatives which are free, so you might point your agent there. I’m sorry for that - we are working our ***** off to get the most juice out of our gpus! Alex

1 0

Minimax 2.5
by Strube, Alexandre 16 Feb '26

16 Feb '26

Hello Blabladores! Sharp-eyed people might have noticed that, during the weekend, a new model landed on Blablador: Minimax 2.5. It’s a substantial upgrade from Minimax 2.1, which was already a good model. It’s also visibly faster. Benchmarks say little, but here we go: One of the things they put some effort on was coding skills. Right from them: "A significant improvement from previous generations is M2.5's ability to think and plan like an architect. The Spec-writing tendency of the model emerged during training: before writing any code, M2.5 actively decomposes and plans the features, structure, and UI design of the project from the perspective of an experienced software architect. M2.5 was trained on over 10 languages (including Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, and Ruby) across more than 200,000 real-world environments. Going far beyond bug-fixing, M2.5 delivers reliable performance across the entire development lifecycle of complex systems: from 0-to-1 system design and environment setup, to 1-to-10 system development, to 10-to-90 feature iteration, and finally 90-to-100 comprehensive code review and system testing. It covers full-stack projects spanning multiple platforms including Web, Android, iOS, and Windows, encompassing server-side APIs, business logic, databases, and more, not just frontend webpage demos. " There are more plots at their announcement, at https://www.minimax.io/news/minimax-m25 One of the things they show is the speed of progress. Minimax M1 was clearly worse than the U.S. closed models by a fair margin. Now, with 2.5, they are roughly on par with the best of the best. Their line is steeper than the others. They talk a lot about how much reinforcement learning improved things, but one thing stands out to me: If you ask minimax what is its name, it will say it’s claude from Anthropic. Seems like some heavy distillation happened here. If I ask: “What’s your name?” it will reply correctly. But if I ask if it’s Claude, it will reply: "Yes, I am Claude, an AI assistant created by Anthropic. How can I help you today?” Anyway, thanks to Konstantin to get this thing running in no time, and let’s bark! Alex, the barkmaster

1 0

alias-huge DOWN for some minutes
by Strube, Alexandre 11 Feb '26

11 Feb '26

Yo, since the migration to the cloud, Minimax and therefore alias-huge were missing one setting which made it never produce the <think> token, so you would see all its thinking on the web ui and in other places. We fixed this, but need to reboot the model. It’s a big one, so it will take a while. It’s happening now. Good night, let’s howl! Alex

1 1

New coding model: Qwen-Next 80b
by Strube, Alexandre 04 Feb '26

04 Feb '26

so far, we had two coding models: Qwen3-Coder-30B-A3B-Instruct, from august. and Devstral-Small-2-24B-Instruct-2512, from december. Qwen3-Coder-Next was released and deployed at blablador yesterday. It’s an 80B parameter model, with only 3b activated at any time - meaning it is FAST In fact, it’s one of our fastest models right now, and a pretty good one. All the three are hooked on alias-code, meaning you might get any of them - which is something I was considering to change. Anyone against me phasing down qwen3-coder from august? By the way, blablador works as a charm with OpenCode and OpenCLAW. I recommend you give it a try - especially opencode. Let’s bark! Alex, sick AF but always working :-)

1 0

GOOD NEWS EVERYONE
by Strube, Alexandre 02 Feb '26

02 Feb '26

Finally, after a long time, we got some new GPUS. This allows us to not need to count with the litany of scripts we’ve been using to get some nodes on the JURECA-WestAI supercomputer. What does that entail: - Stability should to up, a lot: by having the models running on nodes allocated only for Blablador 24/7, availability and stability will go up a lot. No more models failing to launch because the machine is too busy. - Performance improvements: by removing the middle-man of proxy jumps and other dark magic I’ve been doing, things should be smoother. This means more tokens for us! - More models: we want to use the new nodes for the latest and greatest models, but now one can use the available time on JURECA-WestAI to run gigantic models experimentally, suck as Kimi-K2.5, for example. As this would be done in a experimental way as Blablador always did, models will come and go according to needs, requests and voices in my head. I would like to Fritz Niesel and Stefan Kesselheim for making possible to have Blablador to run on the WestAI machine. Without that, the Blablador would have been a much smaller dog. Thanks also to Konstantin Rushschanskii, who has been doing magic with kubernetes and who improved model reliability a lot on the weirdest times of the day. Thanks to Tim Kreuzer for the patience he’s been showing us with our random and absurd questions. Thanks to the sysadmins of Jureca and Julich cloud, especially Sebastian Achilles. You guys are the best. Right now, the only models running on the new hardware are GPT-OSS, a personal favorite of many, and Minimax 2.1, which in my opinion is the best model we have right now. Let’s bark this weekend! Alex, Blablador Honcho

2 2

Helmholtz Codebase outage
by Strube, Alexandre 27 Jan '26

27 Jan '26

Blablador uses Helmholtz Gitlab for authentication, and the server is currently down. Planned Maintenance The Helmholtz Codebase is currently unavailable due to a planned maintenance of the database. Start: January 27th, 15:55 (CET / UTC+1) Planned End: January 27th, 17:00 (CET / UTC+1) Sorry for that :-(

1 0

Do you want an app?
by Strube, Alexandre 23 Jan '26

23 Jan '26

[image] Would you like it? The Siri/android voice integration has been difficult, so it’s basically a website inside an app, like most of them are these days . []s Alexandre Strube surak(a)surak.eti.br

1 0

GLM 4.7 FULL
by Strube, Alexandre 22 Jan '26

22 Jan '26

After the great news of GLM-Flash, I decided to try the full, 385 billion parameter GLM 4.7. And guess what? It’s running! Compared to Minimax, it is slower, but better at math, and multi-modal. It can understand text, video, images and audio. While Minimax is theoretically capable of one million tokens, GLM is limited to 204 thousand. Given that is just slightly better than Minimax, while being slower, I am not sure if I will keep running it. What do you think? Leeeeeeet’s bark! Alex

1 1

What a difference a small patch makes...
by Strube, Alexandre 22 Jan '26

22 Jan '26

Following a hint found on Twitter [1], a little patch on VLLM allowed me to increase the context length on the GLM 4.7-Flash models from 8000 to 200000 (!) tokens. Now, those models get good rankings, but they REALLY overthink. The good thing is that if this patch works with other models, I can run the big GLM 4.7 on bigger nodes! Let’s bark! [1] https://x.com/TheAhmadOsman/status/2013881920099062163

1 1

GLM 4-7 Flash in two variants!
by Strube, Alexandre 20 Jan '26

20 Jan '26

Hello Blabladores! Some of you might have heard of the best medium-sized model released last night, GLM-4.7 Flash. You know the story: It’s slightly better than last week’s model of similar capacity. It’s a 31 billion parameter model, a mixture-of-experts with 3 billion activated. As we hear more and more about strongly quantizing models, I decided to run an experiment. I am running not one, but two versions of GLM 4.7-Flash: One with quantization and one without. Given that they are running on identical hardware (4x RTX 3090 each), something has to give: the not-quantized model has a smaller context length (around 18k tokens, vs 68k tokens for the quantized one). The quantized one is slightly faster, but not by much. I might have to optimize it still. In any case, they are both available for you to compare. What is better: Quantized but bigger context or not-quantized but smaller? You decide. Let’s bark! Alex

1 0