Course contentsShow
← Machine Learning and Deep Learning
Lesson 1626 of 3,53835. Modern Large Language Models: ArchitectureFree lesson

Model Size Classes: 1B, 7B, 13B, 70B+

Common model size tiers in the LLM landscape and their typical capabilities and use cases.

Model Size Classes: 1B, 7B, 13B, 70B+

What you'll learn: You'll understand the common size tiers in the LLM landscape and what capabilities each tier typically offers.

Understanding Model Size Tiers

The LLM ecosystem has naturally crystallized around specific parameter count "classes" that represent different capability-cost trade-offs. These aren't arbitrary—they reflect practical sweet spots discovered through extensive scaling law research and real-world deployment.

The Common Size Classes

1B (Small): Models with ~1 billion parameters are ideal for edge devices, mobile apps, and latency-sensitive applications. They handle basic text completion, simple classification, and straightforward Q&A but struggle with complex reasoning or nuanced instruction-following.

7B (Base): The 7 billion parameter tier represents the minimum size where genuine "smartness" emerges. These models can follow instructions reasonably well, handle multi-turn conversations, and perform basic coding tasks. They're the workhorse tier—small enough to run on consumer GPUs but capable enough for many production use cases.

13B (Enhanced): Doubling to 13 billion parameters brings noticeably improved reasoning, better context understanding, and more reliable instruction-following. This tier offers a good balance for applications needing higher quality without massive infrastructure costs.

70B+ (Frontier): Models with 70+ billion parameters approach state-of-the-art capabilities. They excel at complex reasoning, nuanced language understanding, advanced coding, and multi-step problem-solving. However, they require substantial hardware (multi-GPU setups) and are costly to serve.

Why These Specific Numbers?

These tiers emerged from compute-optimal training principles (like Chinchilla scaling laws) and hardware constraints. For example, 7B fits comfortably on a single A100 GPU with room for batch processing, while 13B requires more memory but remains accessible to many practitioners.

Key Takeaway: Model size classes (1B, 7B, 13B, 70B+) represent practical capability tiers—each roughly doubles the previous, trading increased performance for higher computational and memory requirements.