Guide · 2026-08-31

Which LLM Runs Best on the MacBook Air M4 (16GB, 24GB, or 32GB)?

Last updated 2026-08-31

We tested every MacBook Air M4 RAM tier for local LLM use to find out which models actually run well, and where the fanless design starts to show its limits.

By Mohamed Meguedmi · 10 min read

The MacBook Air M4 in 2026: What Actually Changed for Local LLMs

Apple's M4 finally makes the MacBook Air a legitimate machine for running local language models, not just a compromise pick. The two changes that matter most are memory bandwidth and the RAM ceiling. Memory bandwidth on the M4 is 120GB/s, a 20 percent jump over the M3 and M2 generation, and that difference shows up directly in tokens-per-second once you move past small 7B-8B models into the 14B-and-up range, where the GPU cores spend most of their time waiting on memory rather than compute. The RAM ceiling moved too: for the first time, the Air can be configured with 32GB instead of topping out at 24GB. Combined with the bandwidth increase, that's what lets this machine run a 24B-parameter model in 4-bit quantization comfortably, without constantly swapping to disk. It's a real step up for this line, not a marginal refresh, and it's worth a look at our full MacBook Air M4 buying guide if you're weighing it against the M3 model you might already own.

Under the hood, the M4 in the Air packs a 10-core CPU (4 performance cores, 6 efficiency cores) paired with either an 8-core or 10-core GPU depending on configuration. The Neural Engine has 16 cores rated at 38 TOPS, plenty of headroom for on-device inference, but as of this writing that headroom is only accessible through Apple's own MLX framework and Core ML. Ollama does not yet route inference through the Neural Engine, so most users will still be leaning on the GPU cores for day-to-day chat and coding workloads. Cooling is fully passive, there's no fan in this chassis, which means every LLM workload eventually runs into a thermal ceiling. Apple's efficiency gains with M4 pushed that ceiling out a bit further than on the M3 Air, but they haven't eliminated it.

Spec	MacBook Air M4
Chip	Apple M4 (2024), 10-core CPU (4P+6E)
GPU	8-core or 10-core, config-dependent
Memory bandwidth	120GB/s LPDDR5X (+20% vs M3/M2)
Neural Engine	16 cores, 38 TOPS (MLX/Core ML only, no Ollama support yet)
RAM options	16GB, 24GB, 32GB
Cooling	Fanless (passive)

How Much RAM Do You Actually Need? 16GB vs 24GB vs 32GB

RAM is the single decision that determines which models you can actually run, and on a MacBook Air it's a decision you're stuck with for the life of the machine, Apple's unified memory is soldered in and can't be upgraded after purchase. Our recommendation for most buyers is the 24GB configuration. It lands squarely in the sweet spot for this generation of open-weight models, roughly 8B to 24B parameters at 4-bit to 5-bit quantization, and the $200 step up from the base 16GB model is the best dollar-for-dollar upgrade Apple sells on this machine. The 32GB tier, at $400 over base, only pays for itself if you plan to run 27B-30B models on a regular basis, a real use case, but a distinctly power-user one.

Here's how the usable memory actually breaks down once macOS, background processes, and your browser tabs take their share:

Configuration	Usable for models	What it comfortably runs
16GB (base)	~11-12GB	8B-9B at Q5, Gemma 12B at Q4
24GB (+$200)	~18GB	Qwen 3.5 9B at Q8, Mistral Small 24B at Q4, RAG with a reranker
32GB (+$400)	~25GB	Mistral Small 24B at Q5-Q6, Qwen 3.8 27B at Q4, Qwen3-Coder 30B-A3B, headroom for 32k+ context

The gap between total RAM and usable RAM surprises a lot of first-time buyers. macOS reserves a meaningful chunk for the system and won't let an app claim all of it for the GPU by default, which is why a nominally 16GB machine only comfortably fits an 8B model at higher precision, not a squeezed-in 14B. If you want to model this out for your specific use case before buying, our cost calculator factors in RAM tier against the models you actually want to run.

Which Models Actually Fit Each Configuration

Once you know your usable memory budget, model selection gets straightforward. Here's what we'd point people toward at each tier:

• 16GB (base): Comfortably handles 8B-9B models at Q5 quantization, plus Google's Gemma family around 12B parameters at Q4. That covers roughly 80 percent of what casual local-LLM users actually do, chat, drafting, light coding assistance, and summarization. See our dedicated 16GB configuration guide for a deeper model list.
• 24GB: This is where things open up. You get enough headroom for Qwen 3.5 9B at full Q8 precision (noticeably better output quality than a 4-bit quant of the same model), Mistral Small 24B at Q4, and enough spare memory to run a retrieval-augmented generation pipeline with a reranker model loaded alongside your main LLM.
• 32GB: The ceiling configuration lets you run Mistral Small 24B at Q5-Q6 for better fidelity, Qwen 3.8 27B at Q4, and mixture-of-experts models like Qwen3-Coder 30B-A3B, which only activates a fraction of its parameters per token and runs faster than its total size suggests. You also get real breathing room for extended context windows, 32k tokens and up, without hitting a memory wall.

If you're regularly bouncing between several of these models, our full model catalog is worth cross-referencing before you commit to a RAM tier, since the jump from 16GB to 32GB changes both what you can run and how much budget is left over for storage.

Setting Up Ollama and LM Studio on the M4

Getting a model running on Apple Silicon takes minutes. Ollama is the fastest path if you're comfortable with a terminal:

brew install --cask ollama
brew services start ollama

# Recommended models by RAM tier
ollama run qwen2.5:7b        # 16GB — safe default
ollama run mistral-small     # 24GB
ollama run qwen2.5:32b       # 32GB, watch quantization

LM Studio is the better option if you want a GUI, model browsing, and one-click downloads, and it's also where you'll find Apple's native MLX format alongside the more common GGUF files:

# Download: https://lmstudio.ai/download/mac-arm64
# Then: Discover tab -> search "MLX" for Apple-native builds
# Example: mlx-community/Qwen2.5-9B-Instruct-4bit

The practical difference between the two: Ollama is scriptable and easy to wire into other local tools through its API, while LM Studio's interface makes it easier to compare multiple quantizations of the same model side by side before committing disk space to one. Both tools, and the models they run, are also queryable through the BestLLMfor.com public API (CC BY 4.0) and our open-source MCP server, if you want to pull spec and compatibility data directly into your own scripts or agents instead of looking it up by hand.

M4-Specific Performance Tuning

A handful of settings make a measurable difference on this specific chip, and they're easy to miss if you're following generic Apple Silicon advice written for the M1 or M2 generation.

Favor MLX over GGUF

MLX is Apple's own array framework, built to take advantage of the unified memory architecture directly, and on the M4's higher-bandwidth memory subsystem it pulls further ahead of llama.cpp's GGUF path than it did on M3. In our testing on Granite 4.2 8B, the MLX build ran 8-12 percent faster than the equivalent Q4 GGUF quantization, a gap worth chasing whenever your model of choice ships an MLX build.

Quantize the KV cache on long-context sessions

Recent llama.cpp builds support the -ctk q4_0 -ctv q4_0 flags, which shrink the key-value cache itself, not just the model weights. This is what makes a 32k-token context window realistic even on the base 16GB configuration, where the KV cache would otherwise eat into the same memory pool your model weights need.

Raise the GPU memory ceiling on 32GB machines

By default, macOS caps how much unified memory the GPU can claim (the "wired limit"), and on a 32GB machine the default can leave a 24B model short of what it needs, forcing it to swap. Raising the ceiling reserves more headroom for GPU-addressable memory:

sudo sysctl iogpu.wired_limit_mb=28672
# Make it permanent:
echo "iogpu.wired_limit_mb=28672" | sudo tee -a /etc/sysctl.conf

Consider Core ML for niche, power-efficient workloads

Some quantized models are available in Apple's Core ML format and can run on the 38 TOPS Neural Engine instead of the GPU. It's dramatically quieter and lighter on the battery, but the tooling and model selection here are still thin, this is a path for specific production or on-device deployment use cases, not a general daily-driver setup yet.

MacBook Air M4 vs MacBook Pro M4 vs Mac Mini M4

The Air isn't the only M4 machine worth considering for local inference, and the right pick depends on how you actually work, not just which chip runs fastest in a vacuum.

Machine	Starting price (16GB)	Cooling	Thermal behavior under sustained LLM load
MacBook Air M4	$999	Fanless	Throttles 15-20% after ~10-12 min on 14B+ models
MacBook Pro 14" M4	$1,599	Active fans	No throttling; ~10% higher sustained speed
Mac mini M4	$599	Active fans (desktop)	No throttling; cheapest path to equivalent sustained performance

The rule of thumb we'd give: pick the MacBook Air M4 if portability and silent operation matter to you, it's the machine you can use on a plane or in a coffee shop without a fan spinning up. Pick the Mac mini M4 if you work from a fixed desk, since you're paying $400 less than the Air for equivalent sustained performance once thermals are factored in, an unbeatable value for a stationary setup. Pick the MacBook Pro M4 if you need mobility without any compromise, its active cooling means no throttling during long batch jobs or extended coding sessions, at a $600 premium over the Air. For anyone who expects to push past 30B-parameter models regularly, our MacBook Pro M4 comparison is worth reading, since the Max variants open up a completely different class of local model.

Where the MacBook Air M4 Falls Short

No fan means no noise, but it also means the Air has a hard ceiling that the Pro and the mini don't share.

• Thermal throttling on sustained loads: After roughly 12 minutes of continuous inference on a 14B-or-larger model, generation speed drops 15-20 percent as the chip pulls back to manage heat. For back-and-forth chat this is barely noticeable since you're not generating continuously. For batch jobs, summarizing a folder of documents, running an evaluation suite, it adds up.
• The 32GB ceiling rules out 70B-class models: Even in aggressive quantization, a 70B model needs more unified memory than any MacBook Air configuration offers. If that's your target, you're looking at a MacBook Pro M4 Max with 64GB or more, or a Mac Studio.
• No SD card slot: A small thing, but it matters if your workflow involves pulling photo or video datasets off a card for a multimodal RAG project, you'll need a dongle the Pro doesn't require.

None of these are dealbreakers for the audience this machine is built for. They're reasons to step up to a different M4 machine, not reasons to avoid the Air. If you're still deciding between RAM tiers and machines, our hardware configurator walks through the trade-offs based on the models and context lengths you actually plan to use.