# Installation and pulling models

# On MacOS v26 (on MacBook Pro M5)

Ollama itself
``` bash
curl -fsSL https://ollama.com/install.sh | sh
```




# Models

## In use

Best quality: the Apple Silicon (MLX) build (~18 GB)
```bash
ollama pull qwen3.8:27b-mlx
```
[![](https://storage.googleapis.com/iau-data-dox/uploads/images/gallery/2026-09/scaled-1680-/oXOAZb04ODGS10t1-image-1790323661250.png)](https://storage.googleapis.com/iau-data-dox/uploads/images/gallery/2026-09/oXOAZb04ODGS10t1-image-1790323661250.png)


Fast: MoE daily driver (~23 GB)
```bash
ollama pull qwen3.6:35b-a3b
```
the official qwen3.6:35b-a3b tag is a 23GB Q4_K_M. That's tight on 32 GB, so if memory gets squeezed, pull a smaller community IQ4_XS instead (~17.7GB):
```bash
ollama pull hf.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF:IQ4_XS
```


Small model: for autocomplete
```bash
ollama pull qwen3.5:9b
```
[![](https://storage.googleapis.com/iau-data-dox/uploads/images/gallery/2026-09/scaled-1680-/GDl2U9D6qaqYmLjU-image-1790323625855.png)](https://storage.googleapis.com/iau-data-dox/uploads/images/gallery/2026-09/GDl2U9D6qaqYmLjU-image-1790323625855.png)




## Tested

* good, but too big, calculations are long (40'--4h), almost can not predict workflow pace (timing). Sometimes takes too much time for smaller tasks.
```bash
ollama pull qwen3-coder:30b
```