Model Size Explorer

AI models come in different sizes measured by their parameters. More parameters generally mean more capability, but also more compute and memory requirements.

Size Comparison

Click on any bar to see details. Bar widths use logarithmic scale for visibility.

Example Models

Llama 3.1 7BMistral 7BQwen2 7BGemma 2 7B

Typical Use Cases

  • Quick chatbots and assistants
  • Text summarization
  • Simple code completion
  • Personal devices / edge deployment

Specifications

Parameters7B
Memory Required8GB
Inference SpeedFast

Agentic Potential with Tools

Small models become dramatically more capable when equipped with tools. A 7B model with tools can accomplish tasks that a 405B model without tools cannot.

Coding agentsAutomated workflowsTool-augmented assistants

Used in: Cursor, Windsurf, Claude Code, and other AI-powered dev tools

Example Models

Llama 3.1 70BQwen2 72BMixtral 8x22B

Typical Use Cases

  • Production-grade assistants
  • Complex code generation
  • Document analysis
  • Multi-turn conversations

Specifications

Parameters70B
Memory Required48GB
Inference SpeedMedium

Example Models

Llama 3.1 405BGPT-4 classClaude 3 Opus

Typical Use Cases

  • State-of-the-art reasoning
  • Complex multi-step problems
  • Research and analysis
  • Enterprise applications

Specifications

Parameters405B
Memory Required400GB+
Inference SpeedSlow

At a Glance

7B
7 billion parameters
Memory8GB
SpeedFast
Examples4 models
70B
70 billion parameters
Memory48GB
SpeedMedium
Examples3 models
405B
405 billion parameters
Memory400GB+
SpeedSlow
Examples3 models

Capability Comparison

How different model sizes perform across various tasks.

View:
Capability7B70B405B
Basic Q&A
Simple questions and answers
✓✓
✓✓
Code Generation
Writing and debugging code
~
✓✓
✓✓
Complex Reasoning
Multi-step logical analysis
~
✓✓
Multi-step Math
Advanced mathematical problems
✓✓
Creative Writing
Stories, essays, poetry
✓✓
✓✓
Legend:
✓✓
Excellent
Good
~
Moderate
Limited
✗✗
Poor

Tool Augmentation

The secret to unlocking small model potential

How Tools Work

Tools are functions with descriptions that nudge the model's latent space. They provide structured ways for models to interact with the outside world, compensating for knowledge and capability gaps.

The 5 Agent Primitives

Read
List
Bash
Edit
Search

These 5 primitives enable coding agents to accomplish complex software engineering tasks

Key Insights

  • A 7B model with tools can accomplish tasks a 405B model without tools cannot
  • Mini-SWE-agent (small model) scored 68% on SWE-bench
  • It's just "300 lines of code running in a loop with LLM tokens"

"Less is more" - Context optimization matters more than raw model size. The model does the heavy lifting; tools just give it leverage.

Learn How Agents Work

Discover the agent loop, the five primitives, and what makes models agentic.

Understanding Model Sizes

Parameters are the learned values in a neural network. More parameters allow the model to capture more complex patterns and nuances in language.

VRAM (Video RAM) is GPU memory. Larger models need more VRAM to run. The values shown are approximate for running at 16-bit precision.

Speed vs Capability: Smaller models respond faster but may produce lower quality outputs. The right size depends on your use case and hardware.