Model size is measured in parameters:
- B: Entry-level LLM, fine-tunable on consumer GPUs
- B: Better quality, still accessible
- B: High quality, needs multiple GPUs or quantization
- B+: State-of-the-art, requires data center resources
Scaling laws show predictable improvement with size. But bigger isn't always better for your task. A well-fine-tuned B can beat a generic B model.