Hello, fellow Blabladores! You might have noticed a couple things, and if not, means that everything is working, which is also good! If you haven’t (some things might have appeared in previous emails). Mind you that many of the things here are not properly scaled, so things might be slow, or even disappear. NEW MODELS We have been checking new European models: Aleph Alpha's Kolibri - a German model. Some people like it. TrustLLM - Those are small versions, mid-training, of the TrustLLM project. There are Jülich people involved in it. EVE Instruct: By the Pi-School, associated with the European Space Agency. A model trained on Earth Sciences data. Apertus 1.5, by Swiss AI: It’s a much better version of the previous Apertus-1. alias-ocr: we have a model dedicated only to OCR. It’s LightOnOCR-3-4B, by the french company Light On AI. Other than that, Qwen-Image 2.1 and Flux2-Dev: Image generation. Type a prompt, get an image. Simple. You can steer the model, with things like size and number of steps (these are diffusion models - the more steps, the slower, but better). alias-voice-drama3: Fish Audio’s Drama 3 is a text-to-speech model you can direct. Check the documentation at https://docs.fish.audio/features/text-to-speech/drama-3 - Be careful with this one, it’s a powerful model! alias-video: This is MiniMax H3. It makes video and audio together. Again, careful when using those. Anonymizer: M-size ShinrAI PII 1.3: This is an anonymization model. For example, you give it a text with personal data, it will change this data, and have anonymized data. Swift-Qwen3.8-27b - this is a version of the amazing Qwen 3.8-27, without Qwen’s proclivity to prolixity. This model is way more laconic than the original Qwen. Recently, a new version came out and we already updated it. Muse-Spark-Open-816B-A42B: Muse Spark is a 816B parameter model from Meta. It’s supposedly the best US made open model. It’s way behind the open ones from China, though. alias-qwen3-4b-embeddings, alias-qwen3-4b-reranker, alias-qwen3-4b-reranking, alias-qwen3-8b-reranking: Those are alternative versions of models for embeddings and reranking. alias-clef: This is a decision model, not an LLM. We already spoke about Laya on a previous email. This is similar, just way more capable than Laya. You can play with a website for it at https://test-d.blablador.fz-juelich.de/ And a VERY interesting one: We have our first collaboration with another Helmholtz Center: HZDR, The Dresden-Rossendorf research center, is running a model which appears natively on Blablador! The good thing is that they have the biggest, bestest gpus one can buy right now for this model. Try the model deepseek-v4.1-flash-at-hzdr - it’s incredibly fast! Thanks to all! I hope more centers will come to contribute to make Blablador even better! TECHNICAL CHANGES Watermarking In order to be compliant to the EU AI act, we have done watermarking of our models. Before, this was a simple “This text is generated by Blablador” when you copied the text from the website. Turns out that this is not enough for them anymore, so we introduced a system similar to that being used by all the big providers. I held this as much as I could, but it’s there now. For more details, please check the math behind, here: https://vllm.ai/blog/2026-09-24-watermarking-in-vllm alias-* changes You might have noticed that alias-large is Qwen3.8-Next-Flash, and, from now on, alias-huge came back to be what it was supposed to be: A gigantic, excellent, slow model. MiniMax 2.7 was just not cutting it anymore, so we moved to MiMo-2.6-Pro for alias-huge. I think that’s it. I am probably forgetting something, but that’s enough for a Friday Afternoon. This e-mail has been 100% barked by a human. No AI was involved. All errors are my fault. Let’s bark! Best, Alex