I have been putting the new Mac Mini M6 through some more serious local AI testing, and this time I stepped up from the smaller models to gemma4:12b-it-q4_K_M.
This is a much more useful test than firing one prompt at each machine and calling it a benchmark. I ran repeated generation tests, cold model loading, different prompt sizes and concurrent requests. Both machines used the exact same Ollama model ID and quantisation.
The comparison machine is my M1 Max Mac Studio with 64 GB of unified memory. The Studio has much more memory headroom and a wider memory system, while the new Mini has the advantage of a newer CPU design.
Both Macs were accessed remotely over Tailscale. There was no monitor, keyboard or mouse involved in the testing.
I stopped every other Ollama runner before starting. During the benchmark, ollama ps showed only gemma4:12b-it-q4_K_M on each machine.
This was the first surprise. The new Mac Mini M6 was consistently faster at generating tokens from this exact 12B model.
The Mini was roughly 24% faster in the longer 1,024-token test. That is not a tiny difference that could be dismissed as benchmark noise.
The results are much closer here. The Studio was faster with the medium-sized prompt, while the Mini was faster with the largest prompt in this run.
I would not overstate the prompt numbers. They are useful directional results, but prompt processing can vary depending on caching and the exact shape of the input. The generation results were more consistent across repeated runs.
I also ran multiple requests at the same time. This is closer to what happens when an Ollama server is being used by an automation, a coding assistant and a human at the same time.
The Mini kept its lead under concurrent load. The gap was smaller than in a single request, but the M6 machine still produced more total tokens per second.
The Mini was faster here as well. Both machines were running Ollama remotely through Tailscale, so the benchmark was not affected by a local monitor or desktop session.
That is a serious warning sign. The model runs, and it runs surprisingly quickly, but it does not leave much room for macOS, background services, automation tools or another workload.
The M1 Max Mac Studio has 64 GB of unified memory, so it has a much more comfortable operating envelope even though it was slower at generating tokens in this particular test.
That does not make it the better local AI machine in every situation. The Studio has vastly more memory headroom, and that matters if you want to load larger models, run multiple services or keep a long context available without swapping.
For a single local model, the M6 Mini is much more capable than its size suggests. For a reliable always-on server, I would choose a smaller model than gemma4:12b-it-q4_K_M unless I was happy to accept heavy swap usage.
The practical conclusion is simple:
For comparison, I also ran a smaller Qwen3.5:2b benchmark and a previous Gemma 4 7.5B benchmark.
If anyone wants the benchmark scripts or has a suggestion for another model that can realistically fit inside 16 GB of unified memory, let me know.
This is a much more useful test than firing one prompt at each machine and calling it a benchmark. I ran repeated generation tests, cold model loading, different prompt sizes and concurrent requests. Both machines used the exact same Ollama model ID and quantisation.
Why test gemma4:12b-it-q4_K_M?
The Mac Mini M6 is a tiny 16 GB machine, so the interesting question is not simply whether it can load a larger model. The real question is whether it can run one at a useful speed without the whole machine falling over.The comparison machine is my M1 Max Mac Studio with 64 GB of unified memory. The Studio has much more memory headroom and a wider memory system, while the new Mini has the advantage of a newer CPU design.
Both Macs were accessed remotely over Tailscale. There was no monitor, keyboard or mouse involved in the testing.
The exact model: gemma4:12b-it-q4_K_M
The model used on both machines was gemma4:12b-it-q4_K_M.- 11.9 billion parameters
- Q4_K_M quantisation
- Ollama model ID: 6114515d63c1
- 8.0 GB model size as reported by Ollama
- 8,192-token benchmark context
- Temperature set to 0
- Thinking disabled for a clean generation speed comparison
- GPU acceleration enabled on both Macs
I stopped every other Ollama runner before starting. During the benchmark, ollama ps showed only gemma4:12b-it-q4_K_M on each machine.
gemma4:12b-it-q4_K_M generation results
These are median results from repeated runs, not one lucky measurement.| Test | Mac Mini M6 | M1 Max Mac Studio |
|---|---|---|
| Cold generation, 128 tokens | 30.4 tokens/sec | 24.1 tokens/sec |
| Warm generation, 128 tokens | 31.0 tokens/sec | 24.6 tokens/sec |
| Warm generation, 512 tokens | 29.1 tokens/sec | 23.1 tokens/sec |
| Warm generation, 1,024 tokens | 28.6 tokens/sec | 22.9 tokens/sec |
This was the first surprise. The new Mac Mini M6 was consistently faster at generating tokens from this exact 12B model.
The Mini was roughly 24% faster in the longer 1,024-token test. That is not a tiny difference that could be dismissed as benchmark noise.
Prompt processing with gemma4:12b-it-q4_K_M
Generation is only half of the story. A local model also has to read the prompt, which matters for document analysis, coding assistants, retrieval workflows and long system instructions.| Prompt size | Mac Mini M6 | M1 Max Mac Studio |
|---|---|---|
| 512 words | 4,405 tokens/sec | 4,302 tokens/sec |
| 2,048 words | 19,735 tokens/sec | 21,053 tokens/sec |
| 4,096 words | 19,004 tokens/sec | 14,791 tokens/sec |
The results are much closer here. The Studio was faster with the medium-sized prompt, while the Mini was faster with the largest prompt in this run.
I would not overstate the prompt numbers. They are useful directional results, but prompt processing can vary depending on caching and the exact shape of the input. The generation results were more consistent across repeated runs.
Concurrent requests: can the Mini handle more than one user?
I also ran multiple requests at the same time. This is closer to what happens when an Ollama server is being used by an automation, a coding assistant and a human at the same time.
| Workload | Mac Mini M6 | M1 Max Mac Studio |
|---|---|---|
| Two requests, 256 tokens each | 26.3 aggregate tokens/sec | 21.7 aggregate tokens/sec |
| Four requests, 128 tokens each | 27.4 aggregate tokens/sec | 21.9 aggregate tokens/sec |
The Mini kept its lead under concurrent load. The gap was smaller than in a single request, but the M6 machine still produced more total tokens per second.
Model loading and remote operation
Cold model loading was also measured:| Test | Mac Mini M6 | M1 Max Mac Studio |
|---|---|---|
| Cold model load | 1.62 seconds | 3.16 seconds |
The Mini was faster here as well. Both machines were running Ollama remotely through Tailscale, so the benchmark was not affected by a local monitor or desktop session.
The important part: 16 GB memory pressure
This is where the story becomes more complicated. The Mini completed the entire benchmark, but the 12B model is not a comfortable fit for a 16 GB always-on machine. The Mini started with around 3.3 GB of swap already in use and reached approximately 5.8 GB during the full suite.That is a serious warning sign. The model runs, and it runs surprisingly quickly, but it does not leave much room for macOS, background services, automation tools or another workload.
The M1 Max Mac Studio has 64 GB of unified memory, so it has a much more comfortable operating envelope even though it was slower at generating tokens in this particular test.
My conclusion on gemma4:12b-it-q4_K_M
The new Mac Mini M6 is a little powerhouse. It was faster than the M1 Max Mac Studio at generating tokens from the same 12B Q4_K_M model, and it remained ahead when handling concurrent requests.That does not make it the better local AI machine in every situation. The Studio has vastly more memory headroom, and that matters if you want to load larger models, run multiple services or keep a long context available without swapping.
For a single local model, the M6 Mini is much more capable than its size suggests. For a reliable always-on server, I would choose a smaller model than gemma4:12b-it-q4_K_M unless I was happy to accept heavy swap usage.
The practical conclusion is simple:
- Best generation speed in this test: Mac Mini M6
- Best memory headroom: M1 Max Mac Studio
- Best small, quiet local AI box: Mac Mini M6
- Best choice for larger models and multiple services: M1 Max Mac Studio
For comparison, I also ran a smaller Qwen3.5:2b benchmark and a previous Gemma 4 7.5B benchmark.
If anyone wants the benchmark scripts or has a suggestion for another model that can realistically fit inside 16 GB of unified memory, let me know.