Instruct finetune
Is there a timeline for chat/instruct finetunes on these models?
Unfortunately no timeline to share yet. We're prioritizing and planning that work now.
Any news on an instruct version?
This really needs some effort; there are no capable enough open source translators and also the processor resource needs are out of reach for almost everyone.
If we want some digileap in Finland, we need chat and instruction models now..
This really needs some effort; there are no capable enough open source translators and also the processor resource needs are out of reach for almost everyone.
If we want some digileap in Finland, we need chat and instruction models now..
I would suggest you check out Gemma 3. It's currently probably the best model that is open-weight (that you can actually run on consumer hardware) for Finnish especially. The 27B model at the very least is pretty good at it. 12B model is also not bad, 4B is probably too small and suffers from it. I've somewhat lost hope for anyone training models from scratch to ever match other commercial models like Llama or Gemma/Qwen. (considering the amount of data they are trained on versus models like Viking etc.)
We expect to finally release some Viking chat models soon, but I would moderate your expectations, because these are last generation model only trained on 2 trillion tokens. However, I think we will have some good news coming for Finnish models soon, and potentially other languages as well.
The utter-project/EuroLLM-9B & utter-project/EuroLLM-9B-instruct are very strong fully open source models for European languages, too.
However, if you need translation, our base models work quite well in a few-shot setting, especially the 33B Viking and 34B Poro, and typically outperform other models in our testing, at least at the sentence level.