Will there be QAT models?
#49
by Regrin - opened
Good afternoon!
I remember that the Gemma 3 family released QAT models.
They demonstrated fairly good accuracy with low memory consumption.
Could you please tell me if there will be such models for the Gemma 4?
That would be very nice!
Agreed. The q4_0 QAT models were really good, would love to see it with Gemma 4. A shame Google and OpenAI were the only ones experimenting with QAT, it makes it so much easier than to navigate in the endless jungle of unsloths, bartowskis, mrdachermachers etc.
Yes, QAT was so amazing! I'd love to see them in Gemma 4. Without that, Gemma 4 can absolutely not compete with GPT-OSS 20b on memory efficiency.