How to Choose an Ollama Model: Parameter Count, Quantization, and Hardware Requirements

There are hundreds of models on ollama.com, and each comes in several sizes. This section teaches you to read these labels and choose a model that your computer can run and that fits your needs.

Parameter count: what does B mean?

7b and 14b in a model name indicate its parameter count. b stands for billion.

• More parameters usually mean a smarter model with more knowledge, but it needs more memory and runs more slowly.
• Labels such as e2b and e4b indicate the number of effective parameters. These models use architectures optimized for devices.
• Labels such as 30b-a3b indicate MoE (Mixture of Experts) models: 30 billion parameters in total, but only about 3 billion are active at a time. They need close to as much memory as a 30B model, but run at speeds closer to a 3B model.

Quantization: what are Q4 and Q8?

Each parameter in an original model is usually stored in 16 bits. Quantization compresses parameters into fewer bits, greatly reducing memory requirements at the cost of a small amount of quality.

Select Tags on a model page to see all available versions, such as qwen3:8b-q8_0.

Estimate how much memory you need

A rough formula:

Model size (GB) ≈ parameter count (B) × bits per parameter ÷ 8

For a 4-bit quantized model, add about 1–2 GB for context (KV cache) and runtime overhead:

The most accurate approach is to check the file size listed on the model page, then add a few gigabytes of headroom. The longer the context you set, the more memory you need (see Environment variables).

The most important rule: the model should fit entirely in VRAM. If it exceeds VRAM and some of it is placed in system memory, speed can drop to a fraction of the original. Choose a smaller model that fits entirely on the GPU.

Recommendations by use case

Models change very quickly. The following are popular model families on Ollama when this article was written and are worth trying first. Check ollama.com for the latest versions.

General chat and Chinese

Reasoning (math and logic)

Models with a thinking mode, such as deepseek-r1, qwen3, and gpt-oss, reason before answering. They work well for problems that require multiple steps, but take longer to respond.

Coding

Code-focused models such as qwen2.5-coder and qwen3-coder are suitable for code completion, explanation, and refactoring.

Vision

Models whose pages are labeled vision, such as gemma3, qwen3.5, and llava.

Text embeddings (for RAG)

nomic-embed-text, mxbai-embed-large, and bge-m3 (multilingual, with strong Chinese performance). These models cannot chat; they only convert text into vectors. See Text embeddings and RAG for how to use them.

How to compare models in practice

Ask different models the same question and enable /set verbose to compare speed:

ollama run qwen3:4b
ollama run qwen3:8b
ollama run gemma3:12b

Prepare 5–10 questions you would actually ask (for example, summarizing your work documents or code you often write). This will help you find the right model better than any leaderboard.

Manage downloaded models

ollama list               # List downloaded models and their sizes
ollama rm gemma3:12b      # Delete unneeded models to free up disk space

See Ollama command reference for more commands.

Updated October 12, 2026