Run Qwen3 8B Q4 locally with Ollama on a 16GB Apple Silicon Mac
A conservative Ollama path for local chat and coding experiments using Qwen3 8B Q4_K_M, Metal acceleration, a modest context window, and a repeatable API check.
Evidence status
Ollama officially supports Apple Metal and Qwen3 is available through its model library. This exact Mac, artifact, context, and prompt combination still needs a submitted run.
Requirements
- Apple Silicon Mac
- 16GB unified memory
- macOS supported by the current Ollama release
- At least 12GB free disk space
Stack
- Runtime
- Ollama · Current stable release
- Model artifact
- Qwen3 8B Q4_K_M · GGUF Q4_K_M
Ollama provides the local model process, CLI, and API; Apple GPU acceleration uses Metal.
The LocalAIRun artifact profile recommends 16GB RAM. Keep context modest on a 16GB system.
Manual mode
Step-by-step execution
Use the same safeguards as the Skill
Check the machine first, use only the linked official sources, confirm large downloads and file locations, and do not bypass macOS security. Mark a step complete only after its check passes.
1. Check memory and disk
- •Confirm the Mac uses Apple Silicon and has at least 16GB unified memory.
- •Close memory-heavy applications and keep at least 12GB disk space free.
Check: The system meets the memory and storage requirements.2. Install and start Ollama
Official instructions →- •Install Ollama from the official download or package documented by Ollama.
- •Launch Ollama and confirm the local service responds before downloading a model.
Check: The Ollama CLI opens and the local service is available.3. Download and run Qwen3 8B
- •Use Ollama's Qwen3 8B model entry and start with the default quantized build.
- •Keep the first context test at 8K or lower and avoid loading another model simultaneously.
Check: The model loads without an out-of-memory error and returns a complete answer.4. Verify and report
- •Send the fixed prompt below through the CLI and local chat API.
- •Record Ollama version, model tag, first-token delay, generation speed if available, memory pressure, and any error.
Check: Both the interactive response and API response complete, or the exact failure is captured.
Baseline settings
- context length
- 8192
- temperature
- 0.2
- concurrent models
- 1
Comparable test
Prompt
Write a small TypeScript function that groups hardware variants by product ID, then explain its time complexity.
Seed: 20260816
Did it run on your hardware?
Success and failure reports both improve this recipe's evidence level.