Apple MacBook Air 15 M5 Review - Very powerful, fanless and without competition

[juri]

12W average in idle?? how can that be, lunar lake is only a third of this.
why are you not doing the battery test with idle and video playback any more??

and still no matte option for the screen, so idiotic without any reason.
until then i wont consider an air.

[How tastes differ]

and still no matte option for the screen, so idiotic without any reason.

It's not 12W, look again. Maybe the matte coating on your screen played a prank on you ;)
And I'm so glad that it's glossy (richer colors, sharper text). Mind that while it's glossy, the anti-reflective coating/ability is very good.

[Interesting for reviewers]

This may be interesting for reviewers:

Apple's Energy Model - Deconstructed

In this video, I reverse engineer Apple's Energy Model on the Mac Studio M4 Max. In the process I explain why and how measured DC power can appear up to 3 times higher than reported M4 Max GPU power.

[..]

[only bandwidth = more tg]

As expected, a vid by Alex Ziskind (youtube.com/watch?v=XGe7ldwFLSE), in this case, proves that Apples's claim of

Peak GPU AI compute: over 4× faster

does not apply to running 3rd party LLMs and only the RAM/unified memory bandwidth increase increases the token generation (28% = 1.28 = 153.6 GB/s (M5) / 120 GB/s (M4)). (153.6 GB/s = 128-bit * 9600 MT/s / 1000 / 8)

[only bandwidth = more tg]

Or look up the 153.6 GB/s and 120 GB/s values here: en.wikipedia.org/wiki/Apple_silicon#M-series_SoCs.

[some llama.cpp benchmarks]

First I confirmed that the CPU and GPU scores of my new Air 15 M5 align with what is expected.

Here are some llama.cpp's llama-bench benchmarks:
First with battery saving mode on:

Code Select Expand

/Users/../llama-b8740/llama-bench --no-warmup -m /Users/../Qwen3.5-9B-UD-Q4_K_XL.gguf -p 128 -n 256 -t 1,2,3,4
ggml_metal_device_init: testing tensor API for f16 support
ggml_metal_library_init_from_source: error compiling source
ggml_metal_device_init: - the tensor API is not supported in this environment - disabling
ggml_metal_library_init: using embedded metal library
ggml_metal_library_init: loaded in 0.013 sec
ggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)
ggml_metal_device_init: GPU name:  MTL0
ggml_metal_device_init: GPU family: MTLGPUFamilyApple10  (1010)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)
ggml_metal_device_init: simdgroup reduction  = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory    = true
ggml_metal_device_init: has bfloat            = true
ggml_metal_device_init: has tensor            = false
ggml_metal_device_init: use residency sets    = true
ggml_metal_device_init: use shared buffers    = true
ggml_metal_device_init: recommendedMaxWorkingSetSize  = 12713.12 MB
| model                          |      size |    params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      1 |          pp128 |        110.11 ± 4.18 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      1 |          tg256 |          9.50 ± 0.30 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      2 |          pp128 |        107.58 ± 6.53 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      2 |          tg256 |          9.45 ± 0.08 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      3 |          pp128 |        110.79 ± 1.14 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      3 |          tg256 |          9.29 ± 0.09 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      4 |          pp128 |        110.78 ± 1.42 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      4 |          tg256 |          8.75 ± 0.95 |With the default no battery saving mode:

Code Select Expand

/Users/../llama-b8740/llama-bench --no-warmup -m /Users/../Qwen3.5-9B-UD-Q4_K_XL.gguf -p 128 -n 256 -t 1,2,3,4
[..]
| model                          |      size |    params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      1 |          pp128 |        216.04 ± 8.47 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      1 |          tg256 |        20.17 ± 0.04 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      2 |          pp128 |        209.43 ± 0.35 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      2 |          tg256 |        20.21 ± 0.03 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      3 |          pp128 |        196.55 ± 5.79 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      3 |          tg256 |        18.34 ± 3.22 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      4 |          pp128 |        193.50 ± 2.13 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      4 |          tg256 |        19.57 ± 0.18 |

[some llama.cpp benchmarks]

On battery saving mode (for no battery saving mode, you guessed it, just like above, multiply by ~2):

Code Select Expand

/Users/../llama-b8740/llama-bench --no-warmup -m /Users/../Qwen3.5-9B-UD-Q4_K_XL.gguf -p 64 -n 64 -t 1-12

| model                          |      size |    params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      1 |            pp64 |        107.20 ± 6.39 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      1 |            tg64 |          9.75 ± 0.06 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      2 |            pp64 |        109.29 ± 0.92 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      2 |            tg64 |          9.62 ± 0.01 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      3 |            pp64 |        109.05 ± 0.96 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      3 |            tg64 |          9.47 ± 0.00 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      4 |            pp64 |        108.81 ± 0.81 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      4 |            tg64 |          9.23 ± 0.13 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      5 |            pp64 |        108.59 ± 0.62 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      5 |            tg64 |          9.20 ± 0.01 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      6 |            pp64 |        108.57 ± 1.21 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      6 |            tg64 |          9.12 ± 0.01 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      7 |            pp64 |        108.34 ± 1.24 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      7 |            tg64 |          9.02 ± 0.04 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      8 |            pp64 |        108.07 ± 1.57 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      8 |            tg64 |          8.97 ± 0.01 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      9 |            pp64 |        107.81 ± 0.76 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      9 |            tg64 |          8.76 ± 0.15 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      10 |            pp64 |        107.75 ± 0.58 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      10 |            tg64 |          8.82 ± 0.03 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      11 |            pp64 |        107.64 ± 0.44 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      11 |            tg64 |          8.71 ± 0.03 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      12 |            pp64 |        107.18 ± 0.34 |
| qwen35 9B Q4_K - Medium        |  5.55 GiB |    8.95 B | MTL,BLAS  |      12 |            tg64 |          8.67 ± 0.05 |
This confirms that the number of threads doesn't matter. On my desktop 7800X3D 4 threads gives pretty much the fastest tokens per second.

[Annabel_S]

You got the screen resolution wrong. It is 2880 x 1864.

[QuantumDot]

This laptop may not have PWM, but that is not the only eye strain factor.

There are also:

* FRC
* Temporal dithering
* KSF phosphor filter

[OLED 90 Hz when, APPLE]

* FRC
* Temporal dithering
* KSF phosphor filter

Maybe going OLED can fix it [and not using PWM].

Also, would u bet ur life that you can notice FRC/temporal dithering most of the time?

Also, maybe Apple's implementation of these is made in such a way, that it's less on the eyes, because I can imagine Apple would care and minimize it, if it's a big issue?

[OLED 90 Hz when, APPLE]

Supposedly BetterDisplay and Stillcolor (note: github.com/aiaf/Stillcolor/issues/27) can disable FRC.

[AI: M6 only up to 32GB]

About AI

M6 supports up to 32GB of unified memory to multitask across demanding apps and run LLMs on device for secure and private agentic tasks.

en.wikipedia.org/wiki/Apple_M6

Well, APPLE, problem it 32 GB RAM/unified memory only gives about this many context tokens for this SoTA AI LLM 35B MoE model (MoE -> fast on RAM/unified memory-only devices)[1]:

.. I've come to the conclusion that (1) 32768 is the biggest context I can get away with in an adequately smart model, and (2) it just ain't enough.

(and this is on a 32 GB Mac, so it applies here perfectly)

Calculating available context: 16,384 [KV cache context tokens per GB]*(32 [GB total memory] - 6 to 8 [GB for the OS] - [quant filesize])[2]:
For huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF -> UD-Q4_K_XL.gguf:
~34,400 context tokens = 16,384*(32-7-22.9) (confirms the quote)

34,400 tokens = 25,800 words.
The 3-bit quant may not be usable at all, even at very low context (at 3-bit this model starts to degrade exponentially), but someone said, it works for their tasks to about 60,000 context tokens. The 4-bit Q4_K_XL quant is good for up to about 80,000 to 120,000 tokens, then the performance is going to degrade, depending on your task. So the higher the quenst, the better will it perform at more context.The full, 16-bit, weights model will be best for its native 262,144 tokens context (extensible up to 1,010,000 tokens).

I have the M5 Air and was hoping for a more than 32 GB RAM config, like at least 38-40 GB, preferably a bit more, ideally 48 GB.
In this memory sense, any gaming laptop that has a 6-8 GB VRAM GPU (used for 700-800 bucks) (it's also going to have upgradable RAM) is superior and cheaper, even if new. The additional 8 GB of memory/VRAM of the GPU make all the difference in being able to fit and run that SoTA model at a useful context size.

[1] artificialanalysis.ai/?models=muse-glimmer%2Cqwen3-6-35b-a3b%2Cqwen3-8-27b%2Cg9v3-39a5b%2Cgemma-4-31b%2Cgemma-4-26b-a4b%2Cqwen3-8-27b-non-reasoning&intelligence=artificial-analysis-intelligence-index#artificial-analysis-intelligence-index
(Qwen3.8-27B scores much higher, but due to being a dense model, will run many times slower)
[2] reddit.com/r/Qwen_AI/comments/1vo8pjz/qwen3827b_kv_cache_works_out_to_64_kibtoken_so/