which dell are you using?I just moved from M3 Macbook Pro 14 inch to 15 inch M4 Air (Apple Store-in trade-in)
Reasons:
1. all my screens including large Dell are 60 hz and I don't game on Macs. I use only SDR content and in anyway, content looks great on M4 with RGB profile - I did not notice any difference with Macbook Pro screen
2. 8GB RAM seems limited so I moved to 16GB of RAM. It is a bit more future proof, though I have to say that 8GB on Macbook Pro never bothered me in any way.
3. SSD is 512GB. This was necessity since MBP also had 512GB.
4. I chose MIdnight as it is closest to black in Macbook Airs.
5. CPU and GPU are leading in its class, so I am happy with that.
The main advantage that it is slightly lighter as I carry it all the time. Also while on mobile, 15 inch is more useful than 14 inch and I have to use bigger text *eyesight aging.
Conclusion: m4 is a modern chip and architecture. Actually, M1 is still very good today as well but I just wanted mignight Air
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Macbook Air M4 vs Macbook Pro M1
- Thread starter sweet160
- Start date
- Sort by reaction score
The battery life on my M4 MBA is astounding . . . weeks go-by before I have to remember to re-charge.
I have MBPro M1 16GB and due to accident someone (not me) broke my screen on accident so it now site Beside my Mac Studio M2 MAX running in clamshell mode, no longer to roam. So I have a new MBAir M4 w/32GB arriving This Friday. I'll be able to test my portable work load on the new device.
In terms of what is best, well it depends on YOUR work load.
However my workloads revolve around AI so I asked perplexity how they compare running LLMs - This is not all that I do with AI but I tend to see more comparable results when I ask about LLMs as opposed to other AI related stuff, I'll probably test these results and create more comparable tests, for now here is what Perplexity said;
The MacBook Air M4 with 32GB RAM significantly outperforms the MacBook Pro M1 with 16GB RAM in running local large language models (LLMs), primarily due to the newer chip architecture, more memory, and improved efficiency and throughput in LLM-related workloads.reddit+1
In terms of what is best, well it depends on YOUR work load.
However my workloads revolve around AI so I asked perplexity how they compare running LLMs - This is not all that I do with AI but I tend to see more comparable results when I ask about LLMs as opposed to other AI related stuff, I'll probably test these results and create more comparable tests, for now here is what Perplexity said;
The MacBook Air M4 with 32GB RAM significantly outperforms the MacBook Pro M1 with 16GB RAM in running local large language models (LLMs), primarily due to the newer chip architecture, more memory, and improved efficiency and throughput in LLM-related workloads.reddit+1
LLM Inference Benchmarks
- Actual LLM benchmarks on the MacBook Air M4 32GB show:
- Phi4-mini (3.8B): 34 tokens/sec
- Gemma3 (4B): 35 tokens/sec
- Llama 3.1 (8B): 20 tokens/sec
- Gemma (27B): 6 tokens/sec
- QWQ (32B): 4 tokens/secreddit
- Previous benchmarks for the MacBook Pro/Air M1 16GB (using Llama 8B Q4): 10-13 tokens/sec for similar models.ominousindustries
- Geekbench and other synthetic CPU benchmarks consistently show that the M4 is about 1.5x–2x faster in single-core and multi-core scores than the M1, but the M4’s true edge for LLMs is the combination of higher RAM ceiling and significantly faster neural engine and memory bandwidth.phonearena+1
- The increased RAM in the M4 Air allows it to run larger LLMs locally without swapping, something the M1 16GB cannot do efficiently, especially for models above 13B parameters.towardsdatascience+1
Benchmark Summary Table
| Model/Spec | RAM | LLM speed (tokens/sec) | Max tested LLM size | Notes |
|---|---|---|---|---|
| MacBook Air M4 | 32GB | 34 (3.8B), 20 (8B), 4 (32B) | 32B params | Smooth at 8-14B, usable up to 32B reddit |
| MacBook Pro M1 (16GB) | 16GB | ~10-13 (8B) | 8-13B params | Feels sluggish above 13B ominousindustries+1 |
Conclusion
The MacBook Air M4 with 32GB RAM is clearly better suited for LLM workloads than the M1 Pro/16GB, providing substantial speed increases and the ability to run much larger models without memory issues. The M4 generation’s improved architecture and memory capacity make it the more powerful and practical choice for running or developing with local LLMs, with peer benchmarks and reports strongly supporting thisng this.reddit+2youtube- https://forums.macrumors.com/threads/macbook-air-m4-vs-macbook-pro-m1.2457819/
- https://ominousindustries.com/blogs...eed-test-localllm-on-m1-vs-m2-vs-m2-pro-vs-m3
- https://www.phonearena.com/reviews/apple-macbook-air-m4-vs-macbook-m1-pro-max_id7168
- https://towardsdatascience.com/local-llm-fine-tuning-on-mac-m1-16gb-f59f4f598be7/
- https://www.geeky-gadgets.com/m4-macbook-or-rtx-4060-developer-llm-benchmark-comparison/
- https://www.linkedin.com/pulse/benc...-apple-m4-pro-vs-rtx-3060-dmitry-markov-6vlce
- https://lowendmac.com/2025/i-had-a-m1-pro-macbook-pro-and-went-for-an-m4-macbook-air/
- https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference
- https://forums.macrumors.com/threads/macbook-pro-llm-performance.2441585/
- https://forums.macrumors.com/threads/anyone-buy-the-m4-macbook-air-with-32gb-of-ram.2455487/
Last edited:
Even without the neural engine - M1 generation SOC is missing bfloat16 support which is heavily used in AI for things that don't run on the neural engine.
For large LLMs you probably aren't running on the neural engine either, fire up activity monitor and you'll see the GPU getting smashed.
As i understand it (and if someone knows better please correct me) the neural engine is good for low power lightweight AI/ML but large models, etc. are better run on GPU.
The Apple neural engine really is quite small, at 38 TOPS it is a bit of a joke compared to a large GPU like say, a 5080 with 450 TOPS.
But... the point is that its very efficient and low power, probably great for background stuff like crunching through your photo library indexing it.
For large LLMs you probably aren't running on the neural engine either, fire up activity monitor and you'll see the GPU getting smashed.
As i understand it (and if someone knows better please correct me) the neural engine is good for low power lightweight AI/ML but large models, etc. are better run on GPU.
The Apple neural engine really is quite small, at 38 TOPS it is a bit of a joke compared to a large GPU like say, a 5080 with 450 TOPS.
But... the point is that its very efficient and low power, probably great for background stuff like crunching through your photo library indexing it.
Last edited:
Indeed the neural engine is not being used, yes I know perplexity mentioned it, LOL, I probably should have removed that part of its output, so I went back and yanked that from my post, so as not to confuse anyone. In fact I have never seen any of the AI stuff I do on My Studio or MBs use the neural engine.
I would say yes. My MBA2 was speedier feeling than my M1 MBP I use for work. The 4 is probably a significant jump.
I have MBP w/16GB RAM and a MBA w/32GB RAM - for most day to day things you probably wont notice much of a difference. If you run LLMs you absolutely will notice a difference, for the most part it depends on what you are doing.
The only thing I've seen that does, sometimes, is MacWhisper for dictation and that is primarily for decoding text, at least if you're on anything higher than base M series machine. For awhile they were using it for everything on all machines by default due to a bug. I've also seen some version of Topaz Labs VideoAI use only the neural engine if you have low power mode turned on but yeah for the most part its GPU all day running LLMs.
I snagged my wife's 13" M1 MBP when we got her a new Air last spring to use as my light portable device, and to me, it is starting to feel like it is 5 years old now. It is still perfectly usable for a lot of use cases, but it does feel significantly slower than the M4 series. I'd personally go for an M4 over an M1 or M1 Pro. Maybe an M1 Max over an M4, but just maybe. The M4 chips are quite good.
When running the same LLMs on Ollama on the MBP M1 vs the MBA M4 the air has a toke rate on average of just under twice what the MBP produces, this is based on several LLMs same size and quants, same prompt, etc,... The MBA M4 is is quite a bit slower than my M2 MAX Studio but I can't remember what the comparison looked like. In terms of other stuff on the MBA that I have played with yeah, the MBA is faster than the MBP, for sure.
Yeah for LLMs a M1 Max would be faster though if it were the 24 core model it would probably be close. No doubt in many GPU tasks, a M1 Max system will be faster, though there are outliers. If I remember correctly, Assassin's Creed Shadows ran better on my base M4 Mac mini than my M2 Ultra, for example. I wouldn't be surprised if Blender runs faster on a base M4 compared to an M1 Max since the M1 series did relatively poorly in that software. A Max/Ultra system will export video quite a bit faster. Like I said, maybe an M1 Max over an M4, but I personally would not go for an M1/M1 Pro over an M4 given the choice unless there was maybe another trade off (price, ram, ssd).
