M5 iPad Pro Assessment: An AI and Gaming Upgrade for AI and Video games That Aren’t There But

How do you evaluation an iPad Pro that’s visually an identical to its predecessor and marginally improves upon its efficiency with a spec bump and a few new wi-fi radios?

Let me attempt:

I’ve been testing the brand new M5 iPad Pro since final Thursday. If you’re a cheerful proprietor of an M4 iPad Pro that you just bought final yr, keep like that; there may be nearly no cause so that you can promote your previous mannequin and get an M5-upgraded version. That’s very true if you happen to bought a high-end configuration of the M4 iPad Pro final yr with 16 GB of RAM, since upgrading to a different high-end M5 iPad Pro mannequin will get you…16 GB of RAM once more.

The story is barely completely different for customers coming from older iPad Pro fashions and people on lower-end configurations, however barely. Starting this yr, the 2 base-storage fashions of the iPad Pro are leaping from 8 GB of RAM to 12 GB, which helps make iPadOS 26 multitasking smoother, but it surely’s not a dramatic enchancment, both.

Apple pitches the M5 chip as a “leap” for native AI duties and gaming, and to an extent, that’s true. However, it’s largely true on the Mac, the place – for a wide range of causes I’ll cowl under – there are extra methods to make the most of what the M5 can supply.

In some ways, the M5 iPad Pro is harking back to the M2 iPad Pro, which I reviewed in October 2022: it’s a minor revision to a superb iPad Pro redesign that launched the earlier yr, which set a brand new bar for what we should always count on from a contemporary pill and hybrid laptop – the sort that solely Apple makes nowadays.

For all these causes, the M5 iPad Pro is just not a really thrilling iPad Pro to evaluation, and I might solely suggest this improve to heavy iPad Pro customers who don’t have already got the (nonetheless outstanding) M4 iPad Pro. But there are a few narratives value exploring in regards to the M5 chip on the iPad Pro, which is what I’m going to concentrate on for this evaluation.

The M5 and Local AI Performance

As I discussed above, the M4 and M5 iPad Pros are almost an identical, each externally and from a specs perspective. They look and weigh the identical. The display screen expertise is similar (excellent) Tandem OLED tech launched final yr. Battery life is similar. Cameras are unchanged.

The key distinction is the M5 chip, which is predicated on a third-generation 3nm course of, versus the M4’s second-gen 3nm course of. More particularly, there are new Neural Accelerators for AI constructed into every of the M5’s 10 GPU cores, which is the place Apple’s claims of three.5× quicker AI efficiency in comparison with the M4 comes from. The M5 additionally provides 12 GB of RAM reasonably than 8 GB on the bottom fashions, and a pair of× quicker learn/write speeds on the interior SSD made potential by the adoption of a PCIe 5 storage controller instead of the M4’s PCIe 4 controller. In addition to the M5, the brand new iPad Pro additionally comes with Apple’s new C1X mobile modem and N1 chip for wi-fi radios.

Apple’s daring declare with the M5 chip is that it was designed for intensive AI duties, which ought to outcome within the aforementioned 3.5× efficiency enchancment over the M4, or 5.6× in comparison with an M1 iPad Pro. The firm additionally claims a giant uplift in graphics workloads and rendering duties because of the M5, resembling 6.7× quicker video rendering with ray-tracing than M1 (4.3× quicker than M4) and 4.8× quicker video transcoding than M4.

Because I’m not a video editor, I couldn’t put these video rendering stats to check. But I’ve been enjoying round with native LLMs for the previous few months, so I used to be intrigued to corroborate Apple’s numbers and, importantly, perceive if there are any sensible advantages for energy customers on iPad.

Unfortunately, whereas Apple’s claims sound engaging, and the Neural Accelerators ought to enhance AI duties on a wide range of fronts, resembling token era per second and prefill time (for time-to-first-token evaluations), these enhancements have little to no sensible use on an iPad Pro in comparison with a Mac proper now. And all of it comes right down to the truth that, regardless of higher multitasking and different options in iPadOS 26, there isn’t a robust app ecosystem to make the most of native LLMs on iPad, starting with Apple’s personal fashions.

Before I share my outcomes, let’s analyze Apple’s claims so we will perceive the place they’re coming from.

Apple carried out AI checks on an iPad Pro with 16 GB of RAM and measured time to first token with “a 16K-token prompt using an 8-billion parameter model with 4-bit weights and FP16 activations and prerelease MLX framework”. You’ll discover a few issues right here. For starters, Apple doesn’t say which app they used to check an MLX-optimized LLM on iPad (I assume it was an inside testing app), and notably, they’re not saying which mannequin they really used. But once I learn that line, it rang a bell.

When Apple launched their Foundation fashions final yr, they mentioned in a footnote that they in contrast their mannequin towards a wide range of different LLMs, together with (amongst others) Llama-3-8B-Instruct. Llama 3 was launched in April 2024, and at first, I assumed that Apple repeated their checks with the identical mannequin for the M5 iPad Pro. However, Llama 3 solely provided an 8K context window, and Apple is saying that they examined a 16K context window for the M5. Therefore, my assumption is that Apple benchmarked both Meta-Llama-3.1-8B-Instruct-4bit or Qwen3-8B-MLX-4bit, each of which occur to be obtainable on the MLX Community page on Hugging Face.

Naturally, I needed to check these fashions myself and see if I might have any sensible use instances for them with my iPad Pro workflow. But I instantly ran right into a sequence of issues, for which solely Apple is in charge:

• There is not any official, Apple-made Terminal app for iPad that allows you to set up open supply initiatives like MLX or different native LLM utilities.
• Apple’s MLX framework is unimaginable on macOS; nonetheless, the corporate hasn’t launched an official MLX consumer that AI builders and lovers can use on an iPad to check it.
• I wasn’t capable of finding an MLX-compatible third-party app that additionally confirmed me key metrics resembling tokens per second (tps), time to first token (TTFT), or milliseconds per token (ms/t). All of those instruments exist on macOS as a result of you’ll be able to simply set up software program from the online; none of them might be put in on the iPad, and because of the economics of the platform, third-party builders aren’t making them, both.
• Does anybody know if Apple’s Foundation mannequin has been optimized for MLX? Does anybody know the precise dimension of their newest native mannequin? Is it completely different from 2024? Why isn’t Apple sharing an even bigger, 8B mannequin for gadgets that may run it, just like the M5 iPad Pro? I proceed to have so many questions on Apple’s Foundation fashions, and the corporate isn’t sharing any new particulars. As I keep saying on AppStories, if Apple needs to be taken severely in AI, they should be extra clear about their work. The undeniable fact that I needed to do loads of digging to guess which fashions Apple used for benchmarks of the M5 says loads. In any case, the offline Foundation mannequin can’t be used for benchmarks, so there’s that.

This is the paradox of the M5. Theoretically talking, the brand new Neural Accelerator structure ought to result in notable good points in token era and prefill time that could be appreciated on macOS by builders and AI lovers because of MLX (extra on this under). However, all these enhancements quantity to little or no on iPadOS at present as a result of there is no such thing as a critical app ecosystem for native AI improvement and tinkering on iPad. That ecosystem completely exists on the Mac. On the iPad, we’re left with a handful of non-MLX apps from the App Store, no Terminal, and the untapped potential of the M5.

In case it’s not clear, I’m coming at this from a perspective of disappointment, not anger. I strongly imagine that Apple is making one of the best computer systems for AI proper now – however these computer systems run macOS. And I’m unhappy that, all issues being equal, an iPad Pro with the M5 doesn’t have entry to the breadth of open-source instruments and LLM apps which are obtainable on macOS.

Nevertheless, you already know me: I wasn’t going to let it go.

Testing Local Models with llama.cpp

Although Apple reported their benchmarks with an MLX mannequin, I figured I’d be capable to see some enhancements in inference time and TTFT with fashions written within the GGUF format, too. After much more analysis in a really fragmented App Store ecosystem, I made a decision to check PocketPal AI, an open-source LLM app that may run GGUF fashions because of llama.cpp. It’s not MLX, however I believed, ”Why not?”

For starters, mannequin dimension: I used to be not in a position to load fashions bigger than 8B ones quantized in 4-bit. For instance, I actually needed to attempt the “small” Qwen3-Coder-30B on the M5 iPad Pro, however that wouldn’t load. Oddly sufficient, Qwen3-8B-gemini-2.5-pro-distill didn’t load on the M5, both.

So for my PocketPal checks, I ended up utilizing Qwen2.5-3B-Instruct and Llama-3.1-8B-Instruct-Q4. PocketPal has a “memory lock” function that forces the system to maintain the mannequin in RAM reasonably than utilizing compression or swapping; I enabled it since I figured it’d even be a superb check for the elevated 153GB/s reminiscence bandwidth of the M5, a virtually 30% enchancment over the M4. Additionally, PocketPal has a “Metal-accelerated API” that may be enabled in Settings, which I examined with Qwen2.5-3B-Instruct first at 50 GPU layers, then at 100. Also, for Qwen2.5-3B-Instruct, I first used a 1,100-token immediate, then a 5,400-token one.

For my Llama-3.1-8B-Instruct-This autumn checks, having seen the outcomes from Qwen, I made a decision to solely attempt a protracted immediate (the place I figured I’d see extra good points in prefill time) with PocketPal’s GPU layers set to 100.

Here are the outcomes:

The outcomes are clear, if modest: throughout my checks, the M5 delivered 1.1× to 1.5× enhancements over the M4 – a far cry from Apple’s “3.5x faster AI” advertising and marketing declare. Initially, I assumed that was as a result of I couldn’t discover a third-party iPad app that supported MLX fashions and exhibiting statistics when operating them.

The greatest win was with Llama-3.1-8B-Instruct-This autumn, the place I noticed a 50% soar in tokens per second (from 8 to 12 tps) and a 27% discount in per-token latency. Qwen2.5-3B-Instruct confirmed related good points, significantly with 100% GPU layers enabled, the place I measured 31–44% enhancements relying on immediate dimension. Longer prompts and full GPU acceleration constantly confirmed the M5’s total baseline good points.

Here’s the factor, although: 12 tokens per second is extra responsive than 8, but it surely’s not a transformative leap – and it pales compared to what might be achieved with MLX instruments on macOS. Apple’s 3.5× declare appears to use to very particular workloads, particularly those which were fine-tuned for MLX. But for sensible LLM inference with llama.cpp – the one which you could simply discover on the iPad App Store nowadays – you must count on incremental enhancements, not a revolution, with the M5.

Still, I wasn’t prepared to surrender.

Making a Custom MLX App for iPad

Since I couldn’t confirm Apple’s claims of three.5× quicker LLM inference on the M5 with public apps from the App Store, I commissioned a customized app for this evaluation. I teamed up with Finn Voorhees, who constructed MLXBenchmark, a easy utility that may set up the newest MLX fashions from MLX Community on Hugging Face, permitting you to speak with LLMs and seize real-time stats for every dialog.

I’m going to share my findings under, but it surely was instantly clear to me once I began evaluating the efficiency of MLX throughout the M4 and M5 that one thing wasn’t including up: I saved getting largely the identical numbers between the previous and new iPad Pro, with marginal good points on the M5 that had been far under Apple’s claims.

So I did extra digging, and I noticed one thing I’d initially ignored: Apple benchmarked fashions operating on the M4 and M5 iPad Pros with a pre-release model of MLX that isn’t obtainable to the general public but. This key element was additionally confirmed by one of many co-creators of MLX, who stated that “much faster prefill” (i.e., how time to first token usually goes down) can be made potential by “a future release” of MLX.

Where is that this “future release”, chances are you’ll ask? This is the place the story will get fascinating.

My understanding is that full MLX assist for Neural Accelerators within the M5 can be rolled out later this yr. A brand new department of the MLX open source repository has just added preliminary assist for Neural Accelerators within the M5, and it may be examined on the M5 MacGuide Pro because it has a Terminal, can set up the newest code from the GitHub repository, and may run mlx-lm.

On iPad, it’s rather less clear. In principle, if builders wish to use the brand new Neural Accelerators for low-level calculations, they’ll already use current APIs (CoreML, TensorOps, Metal Performance Shaders) to take action. In observe, in the event that they wish to construct an LLM app that does that with MLX, they must undergo a number of ranges of dependencies and commits for the MLX department that was launched final evening. I do know, as a result of we did with MLXBenchmark. And even then, the outcomes that I bought when testing completely different fashions had been inconclusive and much under a 3.5× enchancment, which makes me ponder whether neural-accelerated MLX inference can work in any respect on an iPad Pro proper now with this early department.

As issues stand now, I witnessed the identical ~1.2× marginal enhancements with MLX on the M5 iPad Pro that I noticed with a llama.cpp backend. Here are the numbers I benchmarked with our personal MLXBenchmark app operating Qwen3-8B-4bit on the M4 and M5 iPad Pros with the experimental, Metal-accelerated MLX department:

As you’ll be able to see, the M5 delivered 1.2× enhancements in token era (12.2 tokens/s vs 10.2 tokens/s) and barely any enchancment in time to first token with shorter prompts. This 20% achieve is almost an identical to my llama.cpp outcomes, suggesting each frameworks are benefiting from the M5’s quicker reminiscence bandwidth and GPU cores.

The story is barely completely different with longer context home windows: longer prompts appear to trace on the M5’s reminiscence bandwidth benefit extra clearly. With a 20,000-token immediate, the M5 decreased prefill time by 16 seconds (9.8% quicker) in comparison with simply 2 seconds with a 12K immediate, which to me suggests the M5’s ~30% bandwidth improve turns into significant at scale. But even at 20K tokens, the 1.22× era velocity was nowhere close to Apple’s 3.5× declare.

Despite us utilizing the newest MLX department with MLXBenchmark for iPad, I didn’t expertise a 3.5x speedup in native inference. I’d like to know if another iPad app developer can determine this out.

I stay up for testing Apple’s upcoming model of MLX absolutely optimized for the M5 iPad Pro, and I’m significantly eager to see how its new structure can enhance the prefill stage and time to first token.

I do know for a undeniable fact that the Neural Accelerator-based efficiency good points are actual: picture era duties in DrawThings (which labored intently with Apple to supply M5 assist at launch) had been 50% quicker on the brand new iPad Pro because of the M5. My solely challenge is that I discover picture era and so-called “AI art” essentially disgusting, and I’m extra eager to mess around with text-based LLMs that I can use for productivity purposes. You gained’t see me cowl AI picture era right here on MacTales.

Today, with the LLM apps from App Store I might discover and the customized one we constructed for the present model of MLX, I might solely get a 1.2× enchancment over the M4 in real-world LLM inference duties.

Unfortunately, even as soon as Apple launches the brand new model of MLX, I’m nervous that it nonetheless gained’t change the truth that there isn’t a robust ecosystem of apps on iPad that may make the most of the M5 for AI duties. Until Apple builds a Terminal app for iPad, permits for side-loading of software program from the online, or fosters a vibrant ecosystem of recent native iPad apps (or the entire above), the M5’s AI good points on iPad will stay largely theoretical and never sensible.

Enough About AI, What About Multitasking?

Right. I hoped to see some fascinating good points enabled by the M5 for windowing and multitasking, however that wasn’t the case.

On the M5 iPad Pro, I might nonetheless solely open 12 home windows directly within the workspace earlier than considered one of them could be jettisoned again to the app switcher. Even then, identical to on the M4 earlier than, I might see some particular efficiency degradation when all 12 home windows had been open directly: issues would get uneven on-screen, body price began lagging in sure Liquid Glass animations – you already know, the same old.

The multitasking and windowing expertise of the M5 iPad Pro is actually the identical because the M4, regardless of the enhancements to the brand new chip and quicker reminiscence. I’m not able to say that Apple has hit a efficiency wall with their new iPadOS windowing engine already, however on the identical time, I’m unsure why Macs with 16 GB of RAM and far older chipsets might maintain much more home windows open directly again within the day.

The M5 and Gaming

I additionally needed to check the M5’s enhancements for high-end gaming on the iPad Pro. However, equally to the AI story, I bumped into limitations attributable to the absence of video games on the iPad App Store in comparison with Steam for Mac in addition to a scarcity of video games that had been up to date to make the most of the M5’s structure.

Theoretically, the M5 helps third-generation ray tracing with 1.5× quicker rendering than the M4 and 6.7× quicker ray-traced rendering than the M1. In observe, I couldn’t discover a single iPad recreation that featured each applied sciences resembling MetalFX upscaling and ray tracing. There are AAA video games which were up to date for Apple silicon and which assist ray tracing, resembling Cyberpunk 2077 and Control; nonetheless, these video games are Mac-only and can’t be put in on iPad. It’s fairly telling, I suppose, that the one demo I noticed for a recreation that supported 120 fps and ray tracing on the M5 iPad Pro was Arknights: Endfield, which is launching in 2026.

As a outcome, I used to be left testing the few AAA video games for iPad that do assist Metal-based upscaling, however which don’t supply Metal-based frame interpolation or ray-traced rendering. In my checks, carried out with the iPad Pro related to my Mac Studio and operating the MetalHUD overlay to watch stats, I noticed the identical efficiency from the M4 and M5, with almost an identical numbers between the 2. I performed each Resident Evil 4 and Assassin’s Creed: Mirage, with the latter set to ‘High’ graphical settings and 100% rendering decision with the MetalHUD overlay. Both video games had been operating at 30 fps – a tough cap imposed by the video games’ builders that didn’t permit me to bump the body price larger even when I needed to.

Interestingly, Resident Evil 4 confirmed that the GPU contained in the M5 was extra highly effective than the M4’s with a rendering time of 8 ms for every body in comparison with the M4’s 28–30 ms rendering time. Essentially, that meant that the M4’s GPU took 3.5× extra time to render every body than the M5. However, this efficiency hole between M4 and M5, whereas spectacular, is meaningless: once more, as a result of how the sport was developed for iPadOS, each iPads delivered an identical body charges. The M5’s GPU was merely sitting idle more often than not, and though it was rendering every body extra shortly than the M4, the sport’s engine didn’t permit me to decide on a better body price. Assassin’s Creed ran the identical throughout the M4 and M5, and neither recreation supported Metal body interpolation or ray tracing.

So as soon as once more, we return to the identical thought: till builders rebuild their iPad video games to faucet into the M5’s capabilities, or till Apple fosters a richer ecosystem of console-quality video games on iPad, Apple silicon’s enhancements for recreation rendering will stay objectively actual, but in addition theoretical since no one is profiting from M-series chips for gaming to their full extent.

This is a bigger story that goes past the scope of this evaluation in the case of Apple and gaming. The firm has created a outstanding piece of silicon that might permit for console-quality gaming with fashionable options like ray tracing and body era on a fanless laptop that has an OLED show with a 120Hz refresh price and is 5mm skinny. But since there aren’t any video games that unlock the true energy of Apple silicon but, Apple has little or no to indicate for it on iPadOS.

Fast Charging, External Displays, and More

Some of my favourite moments when testing the M5 iPad Pro got here from smaller quality-of-life enhancements that had an actual, measurable influence on my each day workflow.

First up, quick charging. New within the iPad Pro this yr, Apple says that the M5 now helps quick charging as much as 50% in half-hour when utilizing a 60W energy adapter or larger. I can verify that that is correct and a unbelievable change if you’re engaged on the iPad Pro all day. I examined quick charging with this 160W UGreen power adapter and a 100W-certified USB4 cable, and I might certainly cost as much as 50% in half-hour with the iPad’s show turned off on the Lock Screen. Although it will have been good to see quick charging as much as 100W like on different fashionable laptops and telephones, I’ll fortunately take a 50% cost in half-hour. It’s greater than sufficient to get me via a number of hours of labor, and I’m glad to see that the function is just not unique to Apple’s new dynamic charger.

Second, the M5 iPad Pro can now drive exterior shows at 4K decision and a 120Hz refresh price with Adaptive Sync. While Apple itself doesn’t make a 4K@120Hz show (yet?), I personal a 27” 4K OLED monitor that refreshes at 240Hz and is G-Sync appropriate, so I used to be eager to see if this transformation could be supported (a) on non-Apple shows and (b) with USB-C cables as a substitute of Thunderbolt ones.

I’m pleased to report that it labored completely in my checks and I didn’t have to vary something in my setup to make the most of quicker refresh charges. As quickly as I related the iPad Pro to my ASUS monitor, the decision stayed at 4K, and the refresh price was instantly bumped to 120Hz, leading to quicker and smoother animations out of the field with my current USB4-certified cable. This is an wonderful change; once I join the M5 iPad Pro to my desk setup, I now not must sacrifice the standard of the iPad’s inside ProMovement show, and I can take pleasure in the identical, clean iPadOS animations on the ASUS monitor as nicely. It feels, successfully, like utilizing a ProMovement show not made by Apple.

For appropriate shows, Apple can be introducing a brand new choice referred to as ‘Adaptive Sync’ that’s their very own tackle variable refresh rate. In case you’re not accustomed to this function from the PC gaming world, it’s designed to cut back latency; it removes arduous refresh price limits and as a substitute makes a related show refresh at a variable refresh price that modifications in actual time primarily based on what you’re doing. In my case, I noticed that with Adaptive Sync enabled, the monitor’s refresh price would keep idle at 81Hz. As quickly as I moved the iPadOS pointer or any window on-screen, the refresh price would immediately shoot again as much as 120Hz.

Since, like I discussed, I’m not enjoying high-end video games on the iPad Pro – not to mention aggressive video games that will profit from VRR – I’ve chosen to maintain Adaptive Sync off and all the time have my exterior show refresh at 120Hz. But if you happen to’re the form of one who performs aggressive video games on iPadOS and needs to have the very best expertise, I’m guessing you’ll be pleased to allow this setting, assuming your monitor permits for it.

As for the opposite hardware-related options within the iPad Pro, I couldn’t observe any significant modifications – for higher or worse – in Bluetooth 6 or Thread (enabled by the brand new N1 chip) or mobile connections primarily based on the C1X modem. The latter is, arguably, a superb factor: the brand new modem didn’t carry out any higher for 5G connections in my space, but it surely additionally didn’t carry out any worse than its Qualcomm predecessor on the M4 iPad Pro. I constantly bought the identical 5G speeds throughout the M4 and M5 iPad Pro in my neighborhood, and I by no means seen any specific hiccups with the brand new modem. It “just worked”, which can be what I’d say in regards to the C1X in my iPhone Air.

An Aspirational Upgrade

In their announcement of the M5 iPad Pro, Apple wrote:

The new iPad Pro options the subsequent era of Apple silicon, with a giant leap in AI efficiency, quicker storage, and the game-changing capabilities of iPadOS 26.

All of that is true, however as I wrap up this evaluation, I wish to be practical right here. Who’s shopping for an iPad Pro for native AI efficiency at present? Who’s shopping for an iPad Pro for gaming? Where are the first-party Apple apps that faucet into this native AI and ray tracing rendering energy to reveal what is feasible with Apple silicon on an iPad Pro?

Ironically, Apple’s advertising and marketing technique for the M5 on the iPad Pro would work loads higher with “one simple trick”: if solely they borrowed extra points of macOS – resembling a Terminal app, the flexibility to put in software program from the online, or the flexibility to run extra Steam video games with a Proton-like compatibility layer – then the M5 iPad Pro’s story could be loads clearer at present.

Instead, the AI “leap” that Apple mentions is true on paper, however pointless in observe if you buy a brand new iPad Pro this week and begin downloading apps and video games from the App Store. You’ll be caught with the identical assortment of few AAA recreation ports and a restricted number of third-party apps to make use of native LLMs with. And even then, all of these apps must obtain their very own third-party fashions (as a result of the Apple Foundation mannequin merely isn’t ok for many duties) with a fragmented ecosystem of fashions operating through llama.cpp or an older model of MLX. It is, fairly frankly, a multitude.

I’ve little doubt that, over time, the M5’s narrative for native AI and MLX will come into focus, particularly on the Mac. But as issues stand at present, except you’re a video editor or actually like quick charging and a 120Hz refresh price on exterior shows, I wouldn’t suggest upgrading to an M5 iPad Pro over your M4 or, arguably, even M2 iPad Pro.

Apple fastened loads of points with iPadOS 26 and its enhancements to windowing, audio recording, file administration, and extra. But if the M5’s focus is on native AI and recreation rendering, a story as previous because the iPad itself rears its head once more: the iPad’s software program must catch up.