* For text generation and reasoning models: Class A contains models under 20B parameters run on a single consumer GPU. Class B (20-60B parameters) runs on cloud GPUs, class C (66B parameters) need multiple cloud GPUs