AI assisted — https://chatgpt.com/share/ed1db514-7c44-48d2-83dc-97b2772466a3
The landscape of Large Language Models (LLMs) is rapidly evolving, with new models emerging regularly. These models come in various sizes, often denoted by the number of parameters they contain, such as 8 billion (8B) or 175 billion (175B). While this technical nomenclature is useful for specialists, it can be challenging for the general public to grasp. To bridge this gap, we propose a more intuitive and accessible method for sizing LLMs, inspired by t-shirt sizes.
The Growing Spectrum of LLM Sizes
LLMs are expanding in size and complexity, offering unprecedented capabilities in natural language understanding and generation. From compact models that can run on mobile devices to colossal models requiring supercomputer infrastructure, the range of LLMs is vast. However, this variety also brings the challenge of understanding and communicating the capabilities and resource needs of these models effectively.
The Need for Sizing LLMs
Sizing LLMs is crucial for several reasons:
- Resource Consumption: Knowing the size of an LLM helps determine the computational resources required to run it. Smaller models may run efficiently on consumer-grade hardware, while larger models necessitate high-end GPUs or specialized cloud infrastructure.
- Usability: Users need to understand whether a model can run on their device. Simplifying the size descriptions can help non-technical users make informed decisions about which models to use for their applications.
Making Sizing Accessible
Technical jargon like “8B” or “32B” can be overwhelming and inaccessible for many. To make sizing more relatable, we propose using t-shirt sizes as a metaphor. This approach categorizes models into intuitive and easy-to-understand groups, such as XS, S, M, L, XL, and so on, making the information more accessible to a broader audience.
The T-Shirt Size for LLMs
Here is a proposal to categorize (and sub-categorize) LLMs based on parameters.
| Category | Parameters | Suitable For |
| ----------------------- | -------------------------------------------- | -------------------------------------------- |
| XS (Extra Small) | Up to 1 billion | Resource-constrained devices |
| S (Small) | 1-5 billion | Mobile devices, e.g. gemma, phi |
| | S1 (Small 1): 1-2 billion | |
| | S2 (Small 2): 2-5 billion | |
| M (Medium) | 5-20 billion | General laptops, e.g. llama3-7b |
| | M1 (Medium 1): 5-7 billion | |
| | M2 (Medium 2): 7-10 billion | |
| | M3 (Medium 3): 10-20 billion | |
| L (Large) | 20-100 billion | Gaming/GenAI laptops, e.g. llama3-70b |
| | L1 (Large 1): 20-30 billion | |
| | L2 (Large 2): 30-50 billion | |
| | L2 (Large 3): 50-100 billion | |
| XL (Extra Large) | 100-500 billion | GPU servers, e.g. llama3-400b |
| | XL1 (Extra Large 1): 100-200 billion | |
| | XL2 (Extra Large 2): 200-500 billion | |
| XXL (Extra Extra Large) | 500 billion-1 trillion | Future AGI models, e.g. gpt-4o |
| | XXL1 (Extra Extra Large 1): 500-700 billion | |
| | XXL2 (Extra Extra Large 2): 700b- 1 trillion | |
| T (Tera) | 1 trillion+ | Future-proof category for the largest models |Benefits of the T-Shirt Size Approach
- Intuitiveness: This approach is relatable and easy to understand, making it accessible to a broader audience.
- Granularity: More detailed categories allow for finer distinctions between models of different sizes.
- Scalability: The system can accommodate new models by adding more subcategories as needed.
- Clarity: Simplifies the communication of model capabilities and resource requirements.
Limitations
While the t-shirt size approach offers many benefits, it has limitations:
- Oversimplification: It may oversimplify the complexity and specific capabilities of models.
- Resource Variability: Not all models with the same number of parameters have the same resource requirements due to differences in architecture and implementation.
- Subjectivity: The boundaries between categories can be subjective and may need periodic revision as models evolve.
Conclusion
The t-shirt size categorization of LLMs offers a user-friendly and scalable approach to understanding and communicating the capabilities of various models. By making model sizes more relatable and intuitive, this method can help users make more informed decisions about which models to use based on their computational resources. As LLMs continue to evolve, this sizing approach provides a flexible framework that can adapt to future developments, making it a practical tool for both technical and non-technical audiences alike.