Is 16GB enough to run an LLM on a MacBook Air?

Yes for small models. A 4-bit model of 4B to 8B parameters needs roughly 3 to 6GB, which leaves room for macOS and a browser. Gemma 4 E4B is the best fit at 16GB in October 2026. Models above about 14B parameters will not fit comfortably alongside other apps.

What is the best app to run local models on a MacBook Air?

LM Studio for a graphical app with MLX support on Apple silicon, Ollama for a command-line tool and a local API that other apps and coding agents can call. Both are free and both run Gemma 4 and Qwen3.8 models.

Quick Answer

Which Local AI Models Can Run on a MacBook Air? (2026)

Q: Which local AI models can run on a MacBook Air?

As of October 2026, a MacBook Air M5 with 16GB runs small models well: Gemma 4 E4B and E2B (128K context, image and audio input) and other models up to about 8B parameters at 4-bit. With 24GB it runs Gemma 4 26B A4B, a mixture-of-experts model that activates about 4B parameters per token and stays fast. With 32GB it can also load dense ~30B models such as Gemma 4 31B and Qwen3.8-27B at 4-bit, but they generate slowly (single-digit tokens per second) because the Air's memory bandwidth is 153GB/s.

Published: October 7, 2026

The short answer

A MacBook Air runs small and mixture-of-experts models well, and dense 30B-class models only slowly. On the M5 Air (16GB standard, 24GB or 32GB optional, 153GB/s memory bandwidth): Gemma 4 E4B on 16GB, Gemma 4 26B A4B on 24GB or more, and Qwen3.8-27B or Gemma 4 31B on 32GB if you can accept single-digit tokens per second. Facts as of October 7, 2026.

What fits, by memory

Air memory	Best models (4-bit)	Approx. weights in RAM	Feels like
16GB	Gemma 4 E4B · Gemma 4 E2B · other 7–8B models	~3–6GB	Fast; good for chat, summaries, private notes
24GB	Gemma 4 26B A4B (MoE) · everything above	~15–17GB	Fast for its quality — only ~4B parameters active per token
32GB	Qwen3.8-27B · Gemma 4 31B (dense) · everything above	~16–19GB	Strong answers, slow output; fine for batch jobs

Leave 6–8GB for macOS and your apps. Weight sizes are our estimates for 4-bit quantised files (about 0.55–0.6GB per billion parameters); the exact download depends on the quantisation you choose.

Why speed falls off on dense models

Generating each token means reading the active weights from memory once, so a rough ceiling is bandwidth ÷ active weight size. At 153GB/s, a dense 27B model at ~16GB tops out around 9 tokens per second before any overhead; Gemma 4 26B A4B reads only ~4B active parameters per token, so it runs several times faster despite a similar file size. That is why the 26B MoE is the sweet spot on an Air. The Air is also fanless: long generations make it throttle sooner than a MacBook Pro.

The models

Gemma 4 E2B / E4B (Google, April 2, 2026; Apache 2.0). Built for on-device use: 128K context, image and video input, and native audio input for speech. The default for a 16GB Air.
Gemma 4 26B A4B (MoE) and Gemma 4 31B (dense). Up to 256K context; Google says the 31B ranked #3 and the 26B #6 among open models on the Arena text leaderboard at launch.
Qwen3.8-27B (Alibaba, open weights since mid-August 2026). A dense vision-language model with thinking control; the strongest coding and agent model that fits a 32GB Air, at the cost of speed.

Larger open models such as Kimi K3 or DeepSeek V4.1 Flash need far more memory than any MacBook Air has; use them through an API (see current API prices).

How to run them

Install LM Studio (graphical, supports Apple’s MLX format) or Ollama (command line plus a local API on port 11434).
Download the model at 4-bit (Q4_K_M in GGUF, or 4-bit MLX).
Start with a 8K–32K context window; long contexts consume memory on top of the weights.
Close heavy apps before loading a 20GB+ model, and keep the Air plugged in for long runs.

Should you buy a MacBook Air for local AI?

Buy 32GB if local models matter at all — it is the only Air configuration that loads the 27B–31B class. Apple’s US store listed MacBook Air M5 configurations from $1,299 on October 7, 2026 (it launched at $1,099 in March and was raised in June). If local models are your main workload, a MacBook Pro with higher bandwidth and active cooling is the better tool.

Best local LLMs for Mac · best small LLMs for a single GPU · what hardware you need to run LLMs locally · Ollama MLX vs Metal on Apple silicon.

Last verified: October 7, 2026.