Until last Friday, my conclusion was like this.
Local LLMs ultimately depend on how large a model you can stably run.
So I cleared out the 3090, went through an RTX 5050 laptop, and eventually arrived at M1 Max 32GB.
In fact, the M1 Max exceeded my expectations.
I was quite surprised to see it load Qwen 27B and read a Flutter project, even passing tests.
At that time, I truly thought
"Model size is more important than GPU name than I thought"
I believed this.
But over the weekend, I did a slightly different experiment.
I had an economic model I created for gaming, and I had it estimate future values based on historical data.
This time, I used Qwen3 14B and 27B on the M1 Max.
Of course, I expected better results, but...
Unexpected results came out.
At first, I mistook the function itself as the starting value, and after correcting it and proceeding with the calculation, it took about 30 minutes to 1 hour to get the final answer.
On the other hand, on the RTX 5050 laptop, I did the same task with Qwen3 9B and 14B, and in about 10 minutes, I almost immediately derived the result in the direction I wanted.
The thought that came to my mind at that moment was just one.
It's NVIDIA after all.
Of course, I'm not saying the M1 Max is bad.
It still has good power efficiency, is quiet, and makes it easy to load large models.
But what matters in actual work is
rather than
"Can you get the desired results in a few minutes?"
seems to be the answer.
By that standard, NVIDIA's CUDA ecosystem and the perceived speed advantage from inference speed was still quite significant.
But while alternating between the M1 Max and RTX 5050 laptop over the weekend, I noticed something else.
There was something more important than inference speed than I thought.
That's model reloading time.
While the M1 Max has the advantage of being able to load large models, situations where new models are read or drop from memory and come back up happened more often than I thought.
The problem is that this process takes anywhere from a few minutes to over 5 minutes.
Especially in agent-type workflows like Roo Code or Continue, situations where context changes, models are switched, or cache hit rates drop keep occurring, and every time that happens, the model starts being reloaded.
Waiting once is fine.
But if this repeats ten, twenty times a day, it's a different story.
In terms of perception, the time spent waiting for model loading felt longer than inference itself.
On the other hand, the RTX 5050 laptop can load smaller models, but the burden when changing or restarting models was much lighter.
Ultimately, in a real development environment
"Can you run 27B?"
rather than
"Can you get an answer right now?"
seemed more important.
For reference, I made the prompts and major parameters used in the test as identical as possible.
I applied major settings like temperature, context, and repeat penalty identically, and compared by only changing the model in each environment.
I also optimized parameters through multiple trials and errors, then ran tests under identical conditions.
Therefore, I believe this result is not simply due to parameter tuning differences, but rather a result of inference speed, model loading time, and cache utilization efficiency in actual operating environments having a greater impact.
So my current conclusion is this.
Value for large models = M1 Max
Quiet development environment = M1 Max
Task completion time = NVIDIA
Coding/inference productivity = NVIDIA still ahead
I went around and came back from 3090 → 5050 → M1 Max,
If asked to choose just one again now, I'd probably go with NVIDIA.
Of course, it's difficult to view this test as an absolute performance comparison.
But at least based on my actual development environment and inference work,
rather than benchmark scores or model sizes
"When does the answer come out?"
was much more important.
In the end, it seems like what remains with local AI is not TPS but productivity.
Conclusion: If you can run the same model, the Laptop 5050 isn't bad either.... that's it
PS) The memory bandwidth difference seems significant too. High-end NVIDIA GPUs are around 1TB/s, but Apple Silicon is M1 100GB/s → Pro 200GB/s → Max 400GB/s → Ultra 800GB/s level, so to get a similar experience, you eventually have to go with Ultra, and from then on the price skyrockets. 😅
