Table of Contents
Apple’s 2025 hardware story can be summarized in one sentence: M4 chips across the lineup, and Apple Intelligence shipping on entry-level products for the first time. The MacBook Air M4 and Mac Studio M4 Max are the core of this wave. Both represent clear spec improvements, but Apple Intelligence’s software integration has left the engineering community somewhat disappointed. This article breaks down what actually matters in this update.
TL;DR
- MacBook Air M4: Baseline RAM jumps to 16GB, M4 chip with 10-core CPU, 12MP camera, price drops to $999 — best value-for-money thin laptop at this price point
- Mac Studio M4 Max: 3.5× faster than M1 Max; M3 Ultra variant can fit models with 600B+ parameters in unified memory
- Apple Intelligence: Writing Tools and system-level summaries work well; Siri cross-app integration still unreliable; developer API access remains limited
MacBook Air M4 (March 2025)
This is the first MacBook Air with an M4 chip, and the first entry-level MacBook to ship with 16GB as the standard configuration. The M4 uses a 10-core CPU (4 performance + 6 efficiency cores), a 25% increase over the M3’s 8-core design, with a 10-core GPU and 16-core Neural Engine.
Key spec changes:
- Base RAM: 8GB → 16GB (same price)
- Front camera: 1080p → 12MP with Center Stage
- Starting price: 13-inch at $999, 15-inch at $1,199 (cheaper than M3 era)
- New color: Sky Blue
On benchmarks, Final Cut Pro rendering of a 4K sequence dropped from 3 minutes 45 seconds on M3 to 2 minutes 58 seconds. For most users upgrading from M3, the difference may not feel dramatic. For M2 and M1 owners, the jump will be significant.
Mac Studio M4 Max (March 2025)
The Mac Studio now comes in two configurations: M4 Max and M3 Ultra.
M4 Max specs:
- 16-core CPU (12 performance + 4 efficiency)
- 40-core GPU
- Up to 128GB unified memory
- 410 GB/s memory bandwidth
M3 Ultra specs:
- 32-core CPU
- 80-core GPU
- Up to 512GB unified memory
- 800 GB/s memory bandwidth
Apple claims the M4 Max is 3.5× faster than the M1 Max version, and the M3 Ultra is 6.1× faster than the highest-spec Intel 27-inch iMac. The chassis is unchanged from the previous generation — the same squat metal cylinder — but if you don’t care about aesthetics, this is the cheapest entry point for workstation-class performance.
Why the Specs Actually Matter
The 16GB Baseline Shift
Moving the MacBook Air baseline from 8GB to 16GB is more significant than it looks. By 2024, 8GB on macOS was already starting to strain under modern workloads: running a local LLM via Ollama would consume the majority of available memory, and combining a few Chrome tabs with a Docker container would push tasks to SSD swap — accelerating SSD wear. 16GB makes the entry-level MacBook Air a genuinely capable machine for daily local AI workloads.
Local LLM Inference Benchmark
The Mac Studio M3 Ultra with 512GB unified memory is currently the most practical consumer machine for running large language models locally. Apple claims it can fit models with over 600 billion parameters in memory — approaching the estimated parameter count of GPT-4. For AI researchers and engineers who need to run large models without cloud dependencies, this spec is meaningful.
The Reality of Apple Intelligence
Apple Intelligence shipped in iOS 18 / macOS Sequoia, and the M4 MacBook Air is the first entry-level Mac to ship with Apple Intelligence out of the box. Here’s an honest assessment:
What works:
- System-level Writing Tools (rewrite, summarize, proofread) available in most text input fields
- Siri can maintain conversational context across follow-up questions
- Image Playground and Genmoji generation is fast and usable
- Priority Notifications meaningfully surfaces high-importance items
What doesn’t work well: Siri’s cross-app integration is the biggest gap. In theory, Siri should handle complex cross-app commands (“Send the latest photo from Instagram to Mom”). In practice, the success rate is around 50%. Voice recognition accuracy for non-English languages — including Chinese — is noticeably worse than ChatGPT.
The developer story: Apple opened third-party app access to on-device foundational models in a late-2025 developer update — but only through limited interfaces like Writing Tools, Image Playground, and Genmoji. The underlying model is not directly exposed. Apple’s strategy is for apps to route data through Siri, positioning Apple as the AI aggregator. Many developers find this frustrating: there’s no way to integrate Apple’s on-device model into a custom AI pipeline.
How It Compares
| MacBook Air M4 | MacBook Pro M4 Pro | Dell XPS 15 (AMD Ryzen AI) | |
|---|---|---|---|
| Starting price | $999 | $1,999 | ~$1,299 |
| Base memory | 16GB | 24GB | 16GB |
| AI accelerator | 16-core Neural Engine | 16-core Neural Engine | AMD NPU |
| Cooling | Fanless | Active cooling | Active cooling |
| Sustained performance | Thermally limited | Unconstrained | Moderate |
The MacBook Air M4’s fanless design means sustained high-load tasks — long LLM inference runs, compiling large projects — will eventually hit thermal throttling. If your workload requires extended full-load performance, the MacBook Pro or Mac Studio is the right choice.
Bottom Line
The M4 MacBook Air achieves “capable AI development machine” at the $999 price point. The 16GB baseline and 10-core CPU make it the most balanced thin laptop at this price. The Mac Studio M4 Max is a strong option for local AI inference workstations.
But Apple Intelligence reminds us that hardware and software are different problems. The M4’s Neural Engine has plenty of compute. Siri’s instability and the limited developer API surface mean Apple’s AI strategy is still chasing Google and OpenAI. The iOS 19 update and next-generation Siri will be the real test.
References
🇺🇸 English
Apple's 2025 hardware update is a story in two acts. Act one: chips. Act two: AI software. The hardware is genuinely impressive. The software? Still a work in progress. Let's dig in.
Start with the MacBook Air M4, which shipped in March. The headline change isn't the chip — it's the baseline RAM. Apple finally moved the entry-level configuration from 8GB to 16GB, at the same price. That matters more than it sounds. By last year, 8GB on macOS was already showing its age. Run a local language model through Ollama, open a few Chrome tabs, throw a Docker container in there — and you're hitting swap memory on the SSD. That means slower performance and faster SSD wear. 16GB makes the $999 MacBook Air a genuinely capable machine for everyday AI workloads.
The M4 chip itself gets a 10-core CPU — four performance cores, six efficiency cores — up from eight cores in the M3. The GPU bumps to ten cores, and there's a 16-core Neural Engine for on-device AI tasks. In real-world rendering tests, a 4K Final Cut Pro sequence that took three minutes 45 seconds on M3 finishes in under three minutes on M4. If you're upgrading from an M2 or M1, you'll feel that jump. M3 users? The difference is more subtle. Apple also swapped the front camera from a 1080p unit to a 12-megapixel sensor with Center Stage, which is a meaningful quality leap for video calls. And the 13-inch starts at $999, the 15-inch at $1,199 — both cheaper than the M3 equivalents were.
Now, the Mac Studio M4 Max. This is a different beast entirely. Sixteen-core CPU, 40-core GPU, up to 128 gigabytes of unified memory, and 410 gigabytes per second of memory bandwidth. Apple puts the performance gain at 3.5 times faster than the M1 Max version — a real jump for anyone still running that machine. There's also an M3 Ultra configuration that goes even further: 32-core CPU, 80-core GPU, up to 512 gigabytes of unified memory at 800 gigabytes per second bandwidth.
Why does memory bandwidth matter so much for AI? When you're running a large language model locally, the bottleneck isn't raw compute — it's how fast you can move model weights between storage and the processor. The M3 Ultra variant in the Mac Studio can reportedly fit language models with over 600 billion parameters entirely in unified memory. That's approaching GPT-4 territory. For researchers and engineers who need to run large models without cloud dependencies, that's a serious capability.
The fanless design of the MacBook Air is worth flagging here. Passive cooling is great for silence and portability, but if your workload runs the chip at full capacity for extended periods — long inference jobs, large compilation tasks — you will hit thermal throttling. The Mac Studio and MacBook Pro don't have this problem. Know your workload before you choose.
Okay, Apple Intelligence. This is where the story gets more complicated.
Apple shipped Apple Intelligence with iOS 18 and macOS Sequoia, and the M4 MacBook Air is the first entry-level Mac to include it out of the box. Some of it works really well. System-level Writing Tools — the ability to rewrite, summarize, or proofread text in almost any input field — are genuinely useful and fast. Siri can now maintain context across follow-up questions in a conversation, which is a real improvement. Priority Notifications does a decent job of surfacing what actually matters. Image generation is fast and functional.
But Siri's cross-app integration is a mess. The pitch is that you can give Siri complex, multi-step commands — something like "grab the latest photo from Instagram and send it to Mom." In practice, that works maybe half the time. The failure rate is high enough that you can't rely on it. And for non-English languages, including Chinese, voice recognition accuracy is noticeably behind ChatGPT. That's a problem for a global product.
The developer angle is frustrating in a different way. Apple opened third-party app access to on-device AI models in a late-2025 update — but only through narrow interfaces: Writing Tools, Image Playground, Genmoji. The underlying model is not directly accessible. Apple wants apps to route everything through Siri, positioning Apple as the AI aggregator across your device. Most developers find that limiting. If you want to build a custom AI pipeline using Apple's on-device model, you currently can't.
So here's where things land. Three takeaways.
First, the 16GB baseline shift is the most practically significant change in this MacBook Air. It's not glamorous, but it's what makes the $999 machine genuinely useful for modern workloads.
Second, the Mac Studio M4 Max — and especially the M3 Ultra configuration — is the most capable consumer hardware for running large models locally. If that's your use case and you want to stay off cloud infrastructure, nothing else at this price point competes.
Third, Apple's AI story is a hardware-software gap problem. The Neural Engine has the compute. Siri doesn't have the reliability, and the developer API doesn't have the openness. iOS 19 and the next-generation Siri will be the real test of whether Apple can close that gap — or whether Google and OpenAI continue to set the pace.
🇹🇼 中文
Apple 今年上半年做了一件很清楚的事：把 M4 晶片塞進每一條產品線，同時讓 Apple Intelligence 從 Pro 機型下放到入門級產品。MacBook Air M4 和 Mac Studio M4 Max 是這波的核心，規格升級是真的，但軟體那邊就有點另一回事了。
先講 MacBook Air。這次最有感的不是 CPU 變快，而是基本款記憶體從 8GB 升到了 16GB，而且售價反而降到 999 美元。這件事比很多人意識到的還重要。8GB 的 Mac 在 2024 年已經開始有點撐不住了，你跑一個本地的小型語言模型，加上幾個瀏覽器分頁和一個 Docker 容器，系統就開始把任務丟到 SSD 上交換，長期下來對 SSD 壽命不太友善。16GB 讓這台入門 MacBook Air 真正有資格被稱為「可以日常跑本地 AI 工作負載的機器」。
CPU 方面，M4 是 10 核心，比 M3 多了兩個效率核，整體提升大概 25%。如果你現在用 M3，升級感可能不強；但如果你還在用 M1 或 M2，這個跨度是會有感的。有一點要記住：Air 是無風扇設計，長時間跑滿載——比如持續跑推論或編譯大型專案——會遇到熱節流，效能會被壓下來。這種情況才需要考慮 MacBook Pro 或 Mac Studio。
說到 Mac Studio，這次有兩個版本，M4 Max 和 M3 Ultra。M4 Max 有 16 核 CPU、40 核 GPU、最高 128GB 統一記憶體，記憶體頻寬達到每秒 410GB。M3 Ultra 更誇張：32 核 CPU、80 核 GPU、最高 512GB 統一記憶體、頻寬直接翻倍到每秒 800GB。Apple 說 M4 Max 比當年的 M1 Max 快了 3.5 倍——這個數字很有說服力。
M3 Ultra 的 512GB 記憶體是整個產品線最值得關注的規格點。Apple 自己說可以在記憶體內跑超過六千億參數的語言模型，這個規模已經接近 GPT-4 的估計參數量了。對於想在本地跑大模型的 AI 研究者或工程師來說，這是目前消費性電腦裡最實際的選項，沒有之一。
然後我們來談 Apple Intelligence，也是這次最令人糾結的部分。M4 MacBook Air 是第一款出廠就能用 Apple Intelligence 的入門級 Mac，所以這個功能現在是主角。
有些地方確實做得不錯：系統層級的寫作工具在大多數輸入框都能用，Siri 現在可以理解上下文接力——你問完天氣，可以直接問「那需要帶傘嗎」，它知道你在說同一件事。Priority Notifications 的排序也挺有用的。
但 Siri 的跨 App 整合是最大的痛點。理論上你應該能說「把 Instagram 上最近那張照片傳給媽媽」然後它就幫你做到，但實際成功率大概五成。中文語音辨識的準確率也比 ChatGPT 差一截。更讓開發者不滿的是，Apple 目前開放給第三方的介面非常有限——你沒辦法直接呼叫裝置端的基礎模型，只能透過 Writing Tools 或 Image Playground 這些包裝好的介面。Apple 想當 AI 聚合器，讓所有資料都過 Siri 這一關，但這個策略讓想自己整合 AI pipeline 的開發者很受限。
好，最後整理三件事：
第一，MacBook Air M4 在 999 美元的價位上，因為 16GB 記憶體這個決定，真的成為了目前同價位最均衡的 AI 開發機。第二，Mac Studio M3 Ultra 的 512GB 記憶體是本地大模型推論目前最強的消費性選項，認真做 AI 研究的人值得認真考慮。第三，硬體和軟體是兩回事，M4 的 Neural Engine 算力夠，但 Apple Intelligence 的 Siri 整合還在補課，2026 年的 iOS 19 和下一代 Siri 才是真正的分水嶺——在那之前，Apple 的 AI 戰略還在追 Google 和 OpenAI 的腳步。
Tags