Ollama model sizes in GB (2026)

Q: How big is a 7B or 8B Ollama model?

About 4.4–5.2 GB at Ollama's default 4-bit quantization: llama3.1:8b is 4.9 GB, qwen3:8b 5.2 GB and mistral:7b 4.4 GB.

Q: How big is a 70B model?

About 42.5 GB: llama3.3:70b is 42.5 GB and deepseek-r1:70b 42.5 GB.

Q: What's the biggest model here?

qwen3:235b at 142.2 GB, followed by gpt-oss:120b at 65.4 GB.

Part of our guides to local LLM disk space · Sizes read from the Ollama registry on 2026-09-29 · Updated every quarter

Short answer: at Ollama's default quantization, small 1–4B models take 0.5–3.3 GB, 7–8B models about 5 GB, 14B models about 9 GB, 27–32B models 17–20 GB and 70B models about 42.5 GB of disk. The eight most popular models together take 114.7 GB.

Download size of every popular model

These are the exact download sizes Ollama reports for each model's default tag (all layers: weights, template and parameters). Families are ordered by total downloads on ollama.com.

Model	Size on disk	Family downloads
`llama3.1:8b`	4.9 GB	120M
`llama3.1:70b`	42.5 GB	120M
`deepseek-r1:1.5b`	1.1 GB	93.3M
`deepseek-r1:7b`	4.7 GB	93.3M
`deepseek-r1:8b`	5.2 GB	93.3M
`deepseek-r1:14b`	9.0 GB	93.3M
`deepseek-r1:32b`	19.9 GB	93.3M
`deepseek-r1:70b`	42.5 GB	93.3M
`nomic-embed-text:latest`	274 MB	87.5M
`llama3.2:1b`	1.3 GB	84.6M
`llama3.2:3b`	2.0 GB	84.6M
`gemma3:1b`	815 MB	40.8M
`gemma3:4b`	3.3 GB	40.8M
`gemma3:12b`	8.1 GB	40.8M
`gemma3:27b`	17.4 GB	40.8M
`qwen3:0.6b`	523 MB	38.3M
`qwen3:1.7b`	1.4 GB	38.3M
`qwen3:4b`	2.5 GB	38.3M
`qwen3:8b`	5.2 GB	38.3M
`qwen3:14b`	9.3 GB	38.3M
`qwen3:30b`	18.6 GB	38.3M
`qwen3:32b`	20.2 GB	38.3M
`qwen3:235b`	142.2 GB	38.3M
`mistral:7b`	4.4 GB	33.7M
`gemma4:12b`	7.6 GB	26M
`gemma4:26b`	18.6 GB	26M
`qwen2.5-coder:1.5b`	986 MB	21.9M
`qwen2.5-coder:7b`	4.7 GB	21.9M
`qwen2.5-coder:14b`	9.0 GB	21.9M
`qwen2.5-coder:32b`	19.9 GB	21.9M
`llava:7b`	4.7 GB	15M
`gpt-oss:20b`	13.8 GB	13.3M
`gpt-oss:120b`	65.4 GB	13.3M
`qwen3-coder:30b`	18.6 GB	9.6M
`phi4:14b`	9.1 GB	7.7M
`llama3.3:70b`	42.5 GB	4.2M
`mistral-small:24b`	14.3 GB	3.1M

Source: Ollama registry manifests (registry.ollama.ai/v2/library/<model>/manifests/<tag>), summed layer sizes, collected 2026-09-29. Download counts from ollama.com. 1 GB = 1,000,000,000 bytes, the same unit Finder and ollama list use.

Check the sizes on your own Mac

ollama list
du -sh ~/.ollama/models

ollama list shows each model's size; du shows what the whole folder really takes, with shared layers counted once.

The same model in two apps takes twice the space

Ollama, LM Studio and Hugging Face / MLX each keep their own copy. The same Qwen2.5 Coder 32B is 19.9 GB in Ollama and 18.4 GB as a 4-bit MLX download, so trying it in both uses 38.3 GB. See how much disk space local LLMs need.

See every model on your Mac at once. Storage Cleaner lists your Ollama, LM Studio and Hugging Face models with their real size, counts shared layers once, and shows when you last used each one, so you know which to remove.

Try Storage Cleaner free

Frequently asked

How big is a 7B or 8B Ollama model? About 4.4–5.2 GB at Ollama's default 4-bit quantization: llama3.1:8b is 4.9 GB, qwen3:8b 5.2 GB and mistral:7b 4.4 GB.

How big is a 70B model? About 42.5 GB: llama3.3:70b is 42.5 GB and deepseek-r1:70b 42.5 GB.

What's the biggest model here? qwen3:235b at 142.2 GB, followed by gpt-oss:120b at 65.4 GB.

Is download size the same as disk space used? Yes, almost exactly: Ollama stores the downloaded layers as they are. Models that share a base (the same weights with a different template) share those layers on disk, so two related models can take less than their sizes added up.

Is download size the same as memory (RAM) needed? No. To run a model you need roughly its size in free memory plus extra for the context window, so a 20 GB model generally needs a Mac with 32 GB or more of memory.

More: how to delete Ollama models · move them to an external drive.