Quick verdict
On an Apple Silicon Mac with 16GB unified memory, start with a 4-bit Qwen3 8B text model and 4096-token context. Use Qwen3 4B if normal applications leave too little memory for the larger model.
What this profile is for
This profile assumes an Apple Silicon Mac with 16GB total unified memory. The model, graphics, macOS, browser, and other apps share that pool; it is not 16GB of dedicated VRAM beside separate system RAM.
Start with one small chat
Follow install LM Studio for a visual setup. In Discover, choose a Qwen3 8B GGUF build labeled Q4_K_M, inspect its file size, and load it with 4096-token context. A different quantization or an MLX package is a different memory case.
For a command-line setup, follow install Ollama, then run ollama run qwen3:8b. In the chat prompt use /set parameter num_ctx 4096 before your first short question. If memory pressure rises sharply, try ollama run qwen3:4b instead. The library's Q4_K_M downloads are about 5.2GB and 2.5GB respectively; file size is only the starting cost.
Worked memory example
Assume a 5.2GB model file, a 2GB planning allowance for runtime and a short context, and 6GB reserved for macOS and your other apps. That is 5.2 + 2 + 6 = 13.2GB, leaving roughly 2.8GB of the 16GB pool. These rounded GB figures illustrate a budget; the 2GB allowance is not a measured requirement for this model.
Open the calculator with Qwen3 8B selected, choose your actual memory and context, and compare its estimate with Activity Monitor's Memory Pressure and Swap Used while chatting. If pressure climbs or swapping grows, shorten context, close another loaded model, or choose the 4B model.
When to try a larger model
Keep the small model if it handles your writing or short coding questions. Increase model size only for a task it fails, and repeat a normal conversation with your usual applications open. Document chat and image input need additional room for their own components.
Gemma 4 12B Unified targets a 16GB memory class, but a 16GB Mac still has to run macOS and your apps. Treat that as a tighter experiment after the smaller model works, rather than the first download.
Recommended next step
Use the Mac setup comparison to choose the app. If loading fails, follow memory troubleshooting before downloading a larger file.