Introduction

As AI language models become more accessible for local deployment, I wanted to test how the new Phi4-Mini from Microsoft performs on my M4 Mac Mini Pro. This base model Mac with 24GB of RAM provides an interesting test environment for running these AI models locally. Let me share my findings comparing Microsoft’s Phi4-Mini against the larger Phi4 model.

Here’s a video of me putting the model through it’s paces:

Setup Details

I’m running Ollama version 0.6.4 (the latest at time of recording) on my M4 Mac Mini Pro. The minimum requirement for running Phi4 models is Ollama version 0.5.13, so we’re well above the threshold.

For testing purposes, I had QuickTime running and capturing my screen, which does consume some GPU resources. This means you might see slightly better performance on your system if you’re not recording simultaneously.

Phi4-Mini: First Impressions

To get detailed metrics during testing, I used the --verbose flag when launching the model, which provides information on generation time and tokens per second.

Test 1: Writing 500 Words

My first request was “Please write 500 words about artificial intelligence.” The results were impressive – 54 Mini completed this task in just 10 seconds. For perspective, larger models like DeepSeek R1 (32 billion parameters) would typically take several minutes for a response of this length.

Test 2: Copyright Handling

I then asked for lyrics to a Beatles song, expecting the model to either struggle or hallucinate content. Instead, Phi4-Mini quickly acknowledged it couldn’t provide copyrighted content – taking just about a second to decline rather than generating inaccurate information. This is actually an excellent response behavior for a model without internet access.

Test 3: Understanding Nuanced Requests

For my third test, I asked: “Generate three prompt examples to use with AI tools for the creation of Python scripts.” The model understood the nuance that I wanted prompt templates, not actual Python code examples – a subtlety that smaller models often miss. It delivered these prompt examples extremely quickly.

Test 4: Public Domain Content

Finally, I asked for Rudyard Kipling’s poem “If” (which is in the public domain), and the model quickly produced an excerpt of the poem.

Comparing with the Full Phi4 Model

For comparison, I ran the same tests on the full 54 model (14 billion parameters) on the same system.

Task	Phi4-Mini	Phi4 (Full)	Comparison
500 words on AI	10 seconds	~40 seconds	4× slower
Beatles lyrics	Correctly declined	Attempted to generate (20 seconds wasted)	54 Mini performed better
Prompt examples	Very fast	~2× longer than Mini	2× slower
Kipling’s “If”	Fast, partial excerpt	26 seconds, complete poem	4× slower but more complete

Performance Insights

What impressed me most was how Phi4-Mini not only delivered responses much faster but also demonstrated strong understanding of nuanced requests. The memory pressure remained manageable throughout testing – well under the 24GB system RAM even when running the larger model.

While the full Phi4 model did provide more complete responses in some cases (like the complete Kipling poem versus an excerpt), the performance difference was substantial – generally 2-4 times faster with the Mini version.

Conclusion

For daily use on an M4 Mac Mini Pro, Phi4-Mini delivers impressive performance, solid reasoning capabilities, and responsible handling of requests for potentially copyrighted content. While the full Phi4 model offers more comprehensive responses in some scenarios, the speed advantage of Phi4-Mini makes it an excellent choice for many local AI applications.

If you’re running AI language models locally on Mac systems, Phi4-Mini strikes an excellent balance between performance and capability – especially if response time is a priority for your workflows.