Text models

---

---

title: Text models slug: text-models

Generally speaking, few universally applicable statements can be made about text models.

However, I'll attempt to outline some patterns:

1. Typically, size alone isn't the deciding factor. For instance, a 70B model from 2023 will likely underperform compared to a 35B model from 2025.
2. Even when file sizes are similar, a 70B model using 4-bit quantization will usually outperform a 35B model with 8-bit quantization.
3. So-called "fine-tuned" models are optimized for specific use cases but often perform worse in other areas compared to the original model. This is evident in models optimized for storytelling that excel at following plotlines but show reduced multilingual capabilities compared to the base model.

Exceptions, as always, prove the rule!