Naming Conventions of LLM Models

Introduction
When we see any LLM model names like GPT-4o, Claude 3 Sonnet, or LLaMA-2-7B-chat we wonder why companies give such weird names to their models. But let me tell you, these names have lots of meanings inside it. They provide lots of information about that model.

• Common Patterns:
  Suffix              Meaning
  Turbo —> Optimised for speed + cost
  Mini —> Smaller + cheaper
  Pro —> High capability
  Flash —> Ultra-fast
  Instruct —> Fine-tuned to follow instructions
  Chat —> Optimised for conversations
  rlhf → trained with human feedback
• Size Hierarchy:
  xxl > xl > large > base > small
• Size Indicators:
  7B, 13B, 70B → parameters
• Versioning:
  v0.1, v1, v2 → iteration of fine-tuning

There are mainly two types of LLMs, lets understand their naming convention one by one.

1. Paid Models

   Paid models are mainly business or customer oriented. So, their naming convention mainly focused on Simplicity, branding and positioning.

   General Pattern in paid models:
   [Model Family] + [Version] + [Variant / Capability Tier]

   Lets see some examples:

   Example 1: GPT-4o
   Breakdown:
   GPT → Model family
   4 → Generation (improvement over GPT-3.5)
   o (omni) → Multimodal capability

   Meaning:
   A 4th-gen model capable of handling text, image, audio, etc.

   Example 2: Gemini 1.5 Pro
   Breakdown:
   Gemini → Model family
   1.5 → Incremental upgrade
   Pro → High capability

   Other variants:
   Flash → Faster, cheaper
   Ultra → Most powerful

   Paid models naming designed for easy understanding of non-technical users, marketing tiers and product differentiation.

2. Open-Source Model

   Open source model naming is more technical and architecture oriented.

   General Pattern in open source models:
   [organization]/[model-family]-[version]-[size]-[variant]-[format]

   Example 1 : meta-llama/Llama-2-7b-chat-hf
   Breakdown:
   meta-llama → Organization
   Llama-2 → Model family + version
   7b → 7 billion parameters
   chat → Fine-tuned for conversation
   hf → Hugging Face format

   Meaning:
   A 7B parameter chat-optimized LLaMA v2 model

   Example 2: mistralai/Mistral-7B-Instruct-v0.1
   Breakdown:
   Mistral-7B → Base model
   Instruct → Instruction-following
   v0.1 → Version of fine-tuning

   Meaning:
   Instruction-tuned version of Mistral 7B (early release)

Final Thoughts:
– Paid models are designed like products
– Open-source models are designed like engineering artifacts

Understanding this difference can help us to select a better model for our specific requirements.

To read more such technical blogs, please follow us on social media. Thanks.