Ollama Models List 2026: 30+ Models Explained With Real Sizes
Which Ollama model should you actually pull? This list covers the ones people really use, with sizes and honest use cases.
💡 What You Will Learn
Which Ollama model should you actually pull? This list covers the ones people really use, with sizes and honest use cases.
📜 Table of Contents
The Model Zoo Problem
Ollama (178,451 stars, 2026-08-14) makes running local LLMs one command, but the model library is huge and the naming (q4_K_M, 7b, 14b) confuses newcomers. The rule of thumb that survives: bigger is smarter, smaller is faster - and your RAM decides your ceiling.
The Models People Actually Use
Llama 3.1 / 3.3 (8B, 70B) - Meta's generalists. The 8B fits 8GB+ RAM and handles most daily tasks; 70B needs ~48GB and is the quality ceiling for local use. Good default.
Qwen 2.5 (7B, 14B, 32B, 72B) - the multilingual all-rounder with strong tool-calling. 7B is the best small model for coding; 32B is the sweet spot on a 24GB GPU. Excellent Chinese and English.
DeepSeek-R1 (7B-671B variants) - the reasoning specialist: shows its chain of thought. Local versions (distilled 7B/14B/32B) give you the 'thinking' behavior on modest hardware - slower but better on logic puzzles.
Gemma 2 / 3 (2B, 9B, 27B) - Google's efficient models. The 2B runs on a phone-class device; 9B is a solid lightweight pick.
Mistral (7B) / Mixtral (8x7B) - the European classic; strong at instruction following.
Phi-4 (14B) - Microsoft's compact reasoning model; punches above its weight on math.
Command R (35B, 104B) - Cohere's enterprise pick with long context and RAG focus.
gpt-oss (20B, 120B) - OpenAI's open-weight line: the 20B is surprisingly strong for its size.
How to See and Manage Your Models
ollama list # installed models
ollama pull llama3.3 # download
ollama run llama3.3 # chat
ollama show llama3.3 --modelfile # config
ollama rm llama3.3 # delete
The RAM Math (approximate)
| Model | Size | RAM needed |
|---|---|---|
| Gemma 2B | 1.6GB | 4GB |
| Qwen 7B | 4.7GB | 8GB |
| Llama 8B | 4.9GB | 8GB |
| Phi-4 14B | 9GB | 16GB |
| Qwen 32B | 19GB | 32GB |
| Llama 70B | 40GB | 64GB |
The Picks by Use Case
- Daily chat: Llama 3.3 8B or Qwen 2.5 7B.
- Coding: Qwen 2.5 Coder 7B/32B.
- Reasoning puzzles: DeepSeek-R1 distilled.
- Small machine: Gemma 2B.
- Best quality on big hardware: Llama 70B or Qwen 72B.
FAQ
How do I know my RAM limit? The model file size must fit in RAM/VRAM; add ~1GB overhead for the runtime.
What does q4_K_M mean? Quantization level - q4 uses ~4 bits per weight, trading a little quality for a much smaller file.
Can I run these on a Mac? Yes - Ollama uses Metal on Apple Silicon; 16GB M-series Macs run 8B models comfortably.
Which model is best for Chinese? Qwen 2.5 - it was trained heavily on Chinese and handles it better than Llama.
