Significant Otter! ❤
This model is indeed Significant Otter (cloaked Gemma 4 model name from arena.ai). I loved it at arena.ai and I spent hours spamming anonymous battles at arena just to get responses from Gemma 4. People at Reddit speculated the coding results these cloaked Gemma 4 models are giving must be from 120B model! And look at that... 26B A4B MoE and 31B dense!!! You've done something otherworldly magical here and I love it! This is a little Gemini at home for when we lose internet connection!
Thank you Google! ❤
Hi @MrDevolver ,
Thank you so much for the kind words, and for spending so much time putting it through its paces on the arena! We had a lot of fun watching the community react to "Significant Otter" and hearing that its coding abilities felt comparable to a 120B model is an incredible compliment to the team.
Delivering that "Little Gemini at Home" experience was exactly what drove the Gemma 4 family. We focused on pushing parameter efficiency with MoE so developers can run powerful, high quality reasoning locally, even offline.
We truly appreciate the enthusiasm and support from the community. We are excited to see what you build with it!
Hi @MrDevolver ,
Thank you so much for the kind words, and for spending so much time putting it through its paces on the arena! We had a lot of fun watching the community react to "Significant Otter" and hearing that its coding abilities felt comparable to a 120B model is an incredible compliment to the team.
Delivering that "Little Gemini at Home" experience was exactly what drove the Gemma 4 family. We focused on pushing parameter efficiency with MoE so developers can run powerful, high quality reasoning locally, even offline.
We truly appreciate the enthusiasm and support from the community. We are excited to see what you build with it!
Thank you very much for response. ❤ Yes, it is very capable, when I finally had a chance to use it on my PC, I was so excited I couldn't even sleep. I spent the whole night playing with it just to see how far can I push it to code some fun games with it and to fix some previously created games which had broken code. The model is so good it feels like having a much bigger model on a small hardware of a home computer. It has some limits, but what is very important to me at least is the code it generates NEVER had any breaking error that would prevent the app from running. There may be some issues here and there, but usually it's not a fundamental flaw that would require rewriting the whole thing from scratch, so the code the model generates mostly feels much more stable and reliable than the code from any other small model of comparable size I've tried so far! 😀