Can I run a serious LLM on my Mac with M-chip?

Yes, and surprisingly well. Mac Studios with M2 Ultra or M3 Max can hold models up to 70B parameters in memory thanks to unified memory. For one user or a small team that is a workable on-prem solution without a GPU server. For multi-user production, pick something else.

Try this first

1. 1Set your RAM budget: 16 GB is enough for 7B (quantised), 32 GB for 13B, 64 GB for 30B, 128 GB+ for 70B. Remember the OS and your work also need RAM.
2. 2Install Ollama or LM Studio. Both run on Apple Silicon with Metal acceleration, no extra config. LM Studio has a GUI, Ollama is CLI with an API server beneath.
3. 3Pick a quantised model: Q4_K_M or Q5_K_M is the right quality-vs-memory compromise for most SMB use. Unquantised 16-bit models need four times the RAM.
4. 4Test latency: a well-tuned M2 Ultra hits 30 to 60 tokens per second on a 13B model. Fine for interactive chat, patience for batch.
5. 5Limits: in practice one Mac means one user at a time. Two concurrent sessions race on memory. For a team pick a Linux GPU server or cloud inference.

None of the above fits?

Describe your situation below. We pass your input plus the steps you already saw to our AI and return tailored next-step advice. If it's too risky to DIY, we'll say so.