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!
