Best Model for Your Air, by RAM

On a MacBook Air, unified memory is everything. The chip generation affects speed (the current M5 Air reads memory at 153 GB/s, an M4 at 120, an M1 at 68), but the amount of RAM decides which models will even fit without forcing macOS to swap to disk — which tanks performance. Find your RAM tier below and start with the matching top pick.

RAMTop pick (index)Also greatSpeed, top pick
8 GBMiniCPM5-2BNanbeige4.2-3B, Gemma 4 E2B, Phi-4 Miniestimated~39 tok/s (M1 Air)
16 GBMiniCPM5-2B (M1–M4 Air); Gemma 4 12B (M5 Air)GPT-oss 20B, Mellum2.1 (coding)estimated~66 tok/s (M4 Air); estimated~17 tok/s (M5 Air)
24–32 GBQwen 3.6-35B-A3BGemma 4 26B-A4B, North Mini Code 1.0estimated~59 tok/s (M5 Air); estimated~50 tok/s (M4 Air)

Picks follow the index’s per-Air pages (14 November 2026): the model that ranks best for that Air on the index’s suitability scale (capability, plus a capped speed bonus) and reaches 15 tok/s there; the 16 GB pick differs by chip because Gemma 4 12B falls under 15 tok/s on the M1–M4. Every Air figure is an estimate from the index’s bandwidth model; per-Air rankings are on the Best by Mac pages.

8 GB Air — small but mighty

An 8 GB Air is the entry point, and it is genuinely useful if you stick to 2B–4B models that leave headroom for macOS. The index’s pick is now MiniCPM5-2B (OpenBMB, Apache 2.0), a 2.5B dense model with a 128K context window, at estimated~39 tok/s on an M1 Air from OpenBMB’s 1.56 GB Q4_K_M file; it has no Ollama library entry, so pull the GGUF from Hugging Face (Step 3). Nanbeige4.2-3B (Apache 2.0), the pick until November, follows at estimated~30 tok/s from a 2.7 GB file. Gemma 4 E2B is the fastest easy option in Ollama, at estimated~44 tok/s on the same Air, and Gemma 4 E4B and Phi-4 Mini trade speed for answer quality (estimated~23 tok/s and estimated~24 tok/s). PrismML’s 1-bit Bonsai 27B squeezes a 27B model into a few gigabytes; its published speeds are for M4 Pro and M5-class chips, so the index has no Air figure for it yet.

16 GB Air — the sweet spot

16 GB is where a MacBook Air becomes a serious local-AI laptop. On an M1–M4 Air the index’s per-Air pick is MiniCPM5-2B again: estimated~66 tok/s on an M4 Air, with most of the memory left free. The index’s 16 GB tier pick, Gemma 4 12B, is the more capable all-rounder, with image input, but as a dense model it runs at only estimated~13 tok/s on an M4 Air, under the 15 tok/s the index asks of a headline pick; on the current M5 Air (153 GB/s) it reaches estimated~17 tok/s and is that Air’s pick. OpenAI’s GPT-oss 20B, a mixture-of-experts model with 3.6B active parameters built to run in 16 GB, sits between them: estimated~29 tok/s on an M2 or M3 Air and estimated~34 tok/s on an M4 Air. Its 14 GB Ollama build is a tight fit, so close heavy apps first. For coding, JetBrains’ Mellum2.1 reads 2.5B parameters per token (estimated~46 tok/s on an M4 Air). More picks are in the 16 GB guide.

24–32 GB Air — punching above its weight

A 24 GB Air, or a 32 GB M5 Air, can run the index’s best pick for a Mac with 32 GB or less: Qwen 3.6-35B-A3B (Apache 2.0, Score 81, tied at the top of the index with MiMo-V2.6-Flash, which is listed first on the alphabetical tie-break and needs a 192 GB Mac Studio). Its mixture-of-experts design reads only ~3B parameters per token, so it runs at estimated~59 tok/s on an M5 Air, estimated~50 tok/s on an M4 Air and estimated~43 tok/s on an M2 or M3 Air. Ollama’s build is 23–24 GB, which is tight on a 24 GB Air: keep the context modest, or take the 32 GB M5 Air. Its dense sibling Qwen3.6-27B (77.2% SWE-bench Verified, vendor-reported) also fits, but at about 6 tok/s estimated on an M4 Air it is too slow for most use; see the 24 and 32 GB guide for smaller alternatives.

No MLX penalty on the Air. A MacBook Air and a MacBook Pro with the same chip run the same tok/s on a fresh run — the GPU, Neural Engine, and MLX framework are identical. The only difference between them is thermal throttling on long workloads, covered next.

Is the MacBook Air Good Enough?

For most people, yes. The thing to understand is the one trade-off Apple made to keep the Air thin and silent: there is no fan. The chip cools passively through the aluminum chassis. That has a clear, predictable effect on local LLMs:

In practice, the Air is excellent for chat, short coding bursts, drafting, and on-the-go private AI. If your workflow is occasional questions rather than hours of nonstop batch generation, the fanless design will rarely get in your way.

How to Run an LLM on Your Air (5 Steps)

Step 1 — Check your MacBook Air's RAM

Click the Apple menu → About This Mac and read the Memory line. That number (8, 16, or 24 GB) is what decides your model. Note it before going further — it maps directly to the table above.

Step 2 — Install Ollama

Ollama is the simplest way to run models on a Mac. Download it from ollama.com, open the .dmg, drag Ollama to Applications, and launch it once so it installs its command-line tool. Then open Terminal and confirm:

ollama --version

You should see a version like ollama version 0.40.16 (current on 14 November 2026); since 0.40.0, Ollama runs models its MLX runtime supports on MLX by default. For a full walkthrough with screenshots, see our Install Ollama on Mac guide.

Step 3 — Pull the right model for your RAM

Run the single command that matches your Air's memory:

# 8 GB Air
ollama run hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
ollama run gemma4:e2b

# 16 GB Air
ollama run gemma4:12b        # M5 Air pick (7.7 GB)
ollama run hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M    # M1–M4 Air pick
ollama run gpt-oss:20b       # reasoning model, a tight fit (14 GB)

# 24–32 GB Air
ollama run qwen3.6:35b-a3b

The first run downloads the model (from about 1.6 GB for MiniCPM5-2B up to 23–24 GB for Qwen 3.6-35B-A3B), then caches it locally so future launches take seconds.

Step 4 — Run and chat

When you see the >>> prompt, you are talking to the model entirely on-device — nothing leaves your Air. Type a question, press Enter, and watch it generate. Type /bye or press Ctrl+D to exit.

Step 5 — Manage heat & throttling

To keep your fanless Air running at its best during longer sessions:

When to Choose a Pro or Studio Instead

The MacBook Air is the right machine for most local-AI users. But if your work is sustained and heavy — running an LLM as a coding agent all day, batch-processing documents, or serving a local API for hours — the fanless design becomes a real limit, and a model that needs more than 24 GB simply won't fit. That is when active cooling and more memory pay off.

Our Mac hardware buying hub breaks down exactly which configuration gives the best tok/s per dollar for local LLMs, across the whole lineup.

Just want the best Air? A MacBook Air M5 with 24 GB of unified memory ($1,499 at Apple) clears the bar for Qwen 3.6-35B-A3B while staying silent and light, and the 32 GB model ($1,699) leaves room for longer contexts. You can grab one here: MacBook Air M5 on Amazon.

As an Amazon Associate, LLMCheck earns from qualifying purchases. This does not affect our rankings.

Bottom line. Match the model to your RAM, run it through Ollama, and the MacBook Air handles real local AI for free and offline. Buy the Air for portability and silence; step up to a Pro or Studio only when you genuinely need sustained, heavy throughput.