+- Mini's Forum (https://forum.minipasila.net)
+-- Forum: AI (https://forum.minipasila.net/forumdisplay.php?fid=3)
+--- Forum: Large Language Models (https://forum.minipasila.net/forumdisplay.php?fid=7)
+--- Thread: DeepSeek V3.1 (/showthread.php?tid=6)
DeepSeek V3.1 - minipasila - 28.08.2025
New model from DeepSeek which is still pretty similar to V3 but now it has hybrid thinking mode so it supposedly can do both non thinking and thinking in the same model but the APIs seem to not know how to enable that (at least Chutes AI).
I'm not sure if it's that much better than the previous model, in terms of multilinguality it seems to be more or less the same. Performs about as well as previous model in Finnish. For RP it seems to use shorter messages overall. And the thinking part appears to cause issues when you use prefills with <tag> like tags. If you have one of those tags in prefill it will spew out random nonsense.
Also considering that GPT-5 is way less censored there's less of a reason to use open-weight models if they are more censored. Though it is still cheaper to use than GPT-5 so price will be a big factor for a while. Even GPT-5 kinda sucks at RP:ing in Finnish so I wonder if I'll ever get a decent Finnish LLM.. SiloAI has done some stuff but it's like several months in the past in the terms of what we have had already. I'm not entirely sure if Gemma 3 is even worse than those Poro 2 models.. They used I think Llama 3.3 70B to generate Finnish data.. even though that model sucks ass in Finnish.. They should have used Gemma 3 27B at the very least because that one is the smallest best model in Finnish at the moment.
But enough about that rant.. DeepSeek V3.1 is an okay model not a huge improvement over the previous one but it's something. Maybe they should add vision to their next model or something or maybe audio... have some competition for GPT-4o..
Links:
https://api-docs.deepseek.com/news/news250821
https://huggingface.co/deepseek-ai/DeepSeek-V3.1
https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base