Apple · M3 · Laptop
Best AI Models for MacBook Air 13" M3 (8 GB)
Runs only small models; 9B doesn't fit in 4.5 GB.
| # | Model | Speed | llmrun Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 22.0 t/s | 22.0 tok/s | Q4_K_M · 3.4 GB | |
| 2 | Gemma 4 E2B ITGoogle Q4_K_M · 3.5 GB · 20.9 t/s | 20.9 tok/s | Q4_K_M · 3.5 GB | |
| 3 | MiniCPM5 2BOpenBMB Q4_K_M · 2.5 GB · 37.7 t/s | 37.7 tok/s | Q4_K_M · 2.5 GB | |
| 4 | Qwen3.5 2BAlibaba Q4_K_M · 1.9 GB · 40.5 t/s | 40.5 tok/s | Q4_K_M · 1.9 GB | |
| 5 | MiniCPM V 4.7 1BOpenBMB Q4_K_M · 1.6 GB · 56.0 t/s | 56.0 tok/s | Q4_K_M · 1.6 GB |
| # | Model | Speed ↓ | llmrun Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 22.0 t/s | 22.0 tok/s | Q4_K_M · 3.4 GB | |
| 2 | MiniCPM V 4.7 1BOpenBMB Q4_K_M · 1.6 GB · 56.0 t/s | 56.0 tok/s | Q4_K_M · 1.6 GB | |
| 3 | Qwen3.5 2BAlibaba Q4_K_M · 1.9 GB · 40.5 t/s | 40.5 tok/s | Q4_K_M · 1.9 GB | |
| 4 | MiniCPM5 2BOpenBMB Q4_K_M · 2.5 GB · 37.7 t/s | 37.7 tok/s | Q4_K_M · 2.5 GB | |
| 5 | Gemma 4 E2B ITGoogle Q4_K_M · 3.5 GB · 20.9 t/s | 20.9 tok/s | Q4_K_M · 3.5 GB |
| # | Model | Speed | Vision Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 22.0 t/s | 22.0 tok/s | Q4_K_M · 3.4 GB | |
| 2 | Qwen3.5 2BAlibaba Q4_K_M · 1.9 GB · 40.5 t/s | 40.5 tok/s | Q4_K_M · 1.9 GB | |
| 3 | MiniCPM V 4.7 1BOpenBMB Q4_K_M · 1.6 GB · 56.0 t/s | 56.0 tok/s | Q4_K_M · 1.6 GB | |
| 4 | LFM3 Omni 2BLiquid AI Q4_K_M · 2.4 GB · 42.2 t/s | 42.2 tok/s | Q4_K_M · 2.4 GB | |
| 5 | Cosmos Reason2 2BNVIDIA Q4_K_M · 3.5 GB · 44.5 t/s | 44.5 tok/s | Q4_K_M · 3.5 GB |
The MacBook Air 13" M3 is the fastest 8 GB machine: it runs only small models. macOS keeps about 3.5 GB, leaving 4.5 GB for models. With 6 TFLOPS it reads long prompts faster than most 8 GB machines: a 4K-token prompt takes 34 s. M3 generation, released 2024; the newer M4 writes -29% faster with 1.2× the compute.
MacBook Air 13" M3 strengths and weaknesses
Weaknesses
- 7–9B models don't fit: Qwen3.5 9B needs 6.6 GB at Q4 vs 4.5 GB usable here
- Slow prompt reading: A 32K-token prompt on Qwen3.5 4B waits 4.5 min for the first token
- No CUDA: vLLM, AWQ, GPTQ and FP8 builds don't run; use GGUF or MLX
Memory for models
4.5 GB of 8 GB9B models need 6.6 GB at Q4, more than the 4.5 GB available. The 16 GB config fits Qwen3.5 9B →Or rent an RTX 3080 Ti (12 GB) from $0.14/hr →
Other memory sizes of the same machine first, then machines and GPUs people weigh against it, each with the one reason it is worth a look. Speeds on Qwen3.5 9B.
Output speed depends on memory bandwidth (102 GB/s). Grey rows need more than the 4.5 GB available.
- 9BQwen3.5 9B · not at 4-bitneeds 6.6 GB
- 14BQwen3 14B · not at 4-bitneeds 11.9 GB
- 27–31BQwen3.8 27B · not at 4-bitneeds 17.9 GB
- 35B MoEQwen3.6 35B A3B · not at 4-bitneeds 22.0 GB
- 70BLlama 3.3 70B Instruct · not at 4-bitneeds 51.3 GB
- 120B MoEGPT OSS 120B · not at 4-bitneeds 70.8 GB
Estimated speeds: Q4_K_M, 4K context, one user. How we estimate
Chip generations: writing vs compute
M1 → M5Writing speed follows memory bandwidth and barely moved: 12.5 to 16.9 tok/s since the M1. Compute, which sets how fast prompts are read, grew 2.8×. Indexed to the M1; speeds on Qwen3.5 9B, same 8 GB.
All 586 models on the MacBook Air 13" M3
Featured onlyBest quant that fits in 4.5 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works
| Model | Quant | VRAM | Speed | Context | Score | Fit |
|---|---|---|---|---|---|---|
Qwen2.5 Coder 1.5BChatCodeFunctions | Q4_K_M | 1.7 GB | 56.0 t/s | 33K | A82, Runs well | |
OLMoE 1B 7B 0924Chat | Q3_K_L | 4.4 GB | 72.3 t/s | 4K | B59, Decent | |
Sweep Next Edit v2 7BChatCode | Q2_K | 4.5 GB | 19.6 t/s | 33K | C30, Usable | |
| Q2_K | 4.1 GB | 62.9 t/s | 128K | A69, Runs well | ||
| IQ2_M | 4.1 GB | 21.3 t/s | 33K | B48, Decent | ||
| IQ3_S | 4.5 GB | 19.6 t/s | 8K | C30, Usable | ||
| Q3_K_M | 3.8 GB | 19.1 t/s | 2K | B55, Decent | ||
Yi 6B ChatChat | Q4_K_M | 4.2 GB | 17.6 t/s | 4K | C38, Usable | |
Qehwa Pashto LlmChat | Q2_K | 4.5 GB | 19.6 t/s | 131K | C30, Usable | |
Saiga Llama3 8BChat | IQ2_M | 4.1 GB | 21.9 t/s | 8K | B47, Decent | |
| IQ2_XXS | 4.4 GB | 28.0 t/s | 33K | C39, Usable | ||
NVIDIA Nemotron Nano 9B v2ChatFunctions | IQ3_S | 4.5 GB | 16.0 t/s | 131K | D24, Barely runs | |
Qwen2.5 7BChatFunctions | IQ3_S | 4.5 GB | 19.6 t/s | 131K | C30, Usable | |
| Q3_K_M | 4.3 GB | 33.2 t/s | — | B49, Decent | ||
Medgemma 4B ITVision | Q3_K_M | 4.5 GB | 31.0 t/s | — | C42, Usable |
Specs
- Chip
- M3
- CPU cores
- 8
- GPU cores
- 8
- Unified memory
- 8 GB
- For models
- 4.5 GB (macOS keeps 3.5 GB)
- Memory bandwidth
- 102.4 GB/s
- FP16 (dense)
- 5.6 TFLOPS
- Neural Engine
- 18 TOPS
- Released
- 2024-03-08
MacBook Air 13" M3 FAQ
How much memory is available for AI on the MacBook Air 13" M3 (8 GB)?
About 4.5 GB of the 8 GB. macOS keeps roughly 3.5 GB for itself and other apps.
What is the largest model the MacBook Air 13" M3 (8 GB) can run?
Qwen3.5 4B (4.7B) is the largest at 4-bit or better: at Q4_K_M it uses 3.4 GB, 76% of the available memory, and runs at about 22.0 tok/s.
Can the MacBook Air 13" M3 (8 GB) run 9B models?
No. Qwen3.5 9B needs 6.6 GB at Q4, more than the 4.5 GB available.
How long does the MacBook Air 13" M3 (8 GB) take to read a long prompt?
On Qwen3.5 4B, about 34 s for a 4K-token prompt and 4.5 min for a 32K-token document before the first word appears.
Is renting a GPU cheaper than buying the MacBook Air 13" M3?
An RTX 3080 Ti (12 GB) rents from $0.10/hr.