The best local LLM for most Macs is Gemma 4: free, quick and able to read images, with the 12B size fitting a 16 GB Mac. On 32 GB, Qwen3.8 27B is the pick for coding and agents; on 24 GB, gpt-oss 20B for step-by-step reasoning; on 64 GB, Llama 3.3 70B; and on 128 GB, gpt-oss 120B. All of them run with nothing sent to a vendor, in free apps such as Ollama and LM Studio. Memory decides what you can run, so start with how much your Mac has.
Key facts
- Everyday work, 16 GB: Gemma 4 12B, 7.7 GB download on a Mac.
- Reasoning, 24 GB: gpt-oss 20B, 14 GB.
- Coding and agents, 32 GB: Qwen3.8 27B, 18.2 GB.
- Big models: Llama 3.3 70B (43 GB) on 64 GB; gpt-oss 120B (65 GB) on 128 GB.
- Cost: nothing. Check each licence before commercial use: Apache 2.0 allows it, Meta's Llama licences add conditions.
On this page
- Best local LLM by Mac memory
- Best local LLM by job
- Which app to run it in
- What local models still can't do
- Frequently asked questions
Best local LLM by Mac memory
| Mac memory | Install first | Download | Licence | Command |
|---|---|---|---|---|
| 8 GB | Llama 3.2 3B | 2.0 GB | Llama 3.2 Community | ollama run llama3.2 |
| 16 GB | Gemma 4 12B | 7.7 GB | Apache 2.0 | ollama run gemma4:12b |
| 24 GB | gpt-oss 20B | 14 GB | Apache 2.0 | ollama run gpt-oss:20b |
| 32 GB | Qwen3.8 27B | 18.2 GB | Apache 2.0 | ollama run qwen3.8:27b |
| 48 GB | Qwen3.6 35B-A3B | 23.1 GB | Apache 2.0 | ollama run qwen3.6:35b-a3b |
| 64 GB | Llama 3.3 70B | 43 GB | Llama 3.3 Community | ollama run llama3.3 |
| 128 GB | gpt-oss 120B | 65 GB | Apache 2.0 | ollama run gpt-oss:120b |
Sizes are what Ollama downloads on a Mac, where it now uses Apple's MLX build of a model if one exists. Each Mac size links to its own page, with more models that fit and my test results. Leave room for macOS, your apps and the conversation itself. To check a model against your exact Mac, use the local LLM memory calculator. Still choosing a Mac? See which Mac to buy for local AI.
Best local LLM by job
- Writing, summaries and everyday questions: Gemma 4 (12B on 16 GB, 26B or 31B on 32 GB).
- Coding: Qwen3.8 27B on 32 GB, or qwen3-coder 30B (19 GB). For a coding agent on your own model, see Claude Code with Ollama.
- Reasoning, maths and planning: gpt-oss 20B on 24 GB, gpt-oss 120B on 128 GB.
- Reading images, screenshots and charts: Gemma 4 on any size, or Qwen3-VL if images are the main job.
- Scanned documents and PDFs: glm-ocr (2.7 GB), a small model built for reading scans.
- Searching your own files: pair a chat model with an embedding model such as nomic-embed-text (274 MB) or bge-m3 for non-English text, in an app like AnythingLLM.
- Agents that call tools: choose a model with the Tools label. See agent skills on local models.
Older models with huge download counts, such as Llama 3.1, DeepSeek-R1 and Gemma 3, still top the charts, but newer ones do the same jobs better. The full table is in the Ollama models list.
Which app to run it in
Ollama runs models in the background for other apps and agents to use. LM Studio gives you a chat window the moment it opens. Both are free. If you can't decide, read Ollama vs LM Studio; many Mac users install both.
What local models still can't do
Local models trail the biggest cloud models on hard reasoning and long, many-step coding jobs. They handle drafting, summaries, document work and most small-business admin well. If you need the top cloud models for some tasks, keep a cloud plan for those and run the rest locally. For prices, see the UK AI price tracker.
Frequently asked questions
What is the best local LLM for a Mac?
For most Macs, Gemma 4: the 12B size on 16 GB, or the 26B or 31B on 32 GB. For coding and agents on 32 GB, Qwen3.8 27B.
What is the best local LLM for a 16 GB Mac?
Gemma 4 12B, a 7.7 GB download on a Mac that reads images and fits with room to spare. Install it with ollama run gemma4:12b. Gemma 4 E4B answers faster, at 87 tokens a second in my tests, but on a Mac it is a 9.5 GB download and a tight fit.
Can a Mac run a 70B model?
Yes, with 64 GB of memory. Llama 3.3 70B is a 43 GB download at 4-bit.
Are local LLMs free?
Yes, to download and run. The licence still sets the rules: Apache 2.0 allows commercial use, while Meta's Llama licences add conditions.
Sources
- Ollama models list, read from ollama.com on 25 October 2026
- Ollama, gemma4 (read 21 November 2026)
New models arrive every week. See Local AI This Week for the latest. Spotted something out of date? Tell me.



