Apple · M3 · Laptop
Best AI Models for MacBook Air 15" M3 (24 GB)
Runs 27–31B models at 4 tok/s; 9B and 14B fit but are slow (11 and 8 tok/s), and 70B doesn't fit in 20.5 GB.
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Q4_K_M · 16.5 GB | 22.6 tok/s | |
| 2 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB | 11.3 tok/s | |
| 3 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB | 24.4 tok/s | |
| 4 | Gemma 3 4B ITGoogle Q4_K_M · 5.0 GB | 25.2 tok/s | |
| 5 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB | 22.0 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Q4_K_M · 16.5 GB | 22.6 tok/s | |
| 2 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB | 24.4 tok/s | |
| 3 | Gemma 3 4B ITGoogle Q4_K_M · 5.0 GB | 25.2 tok/s | overall |
| 4 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB | 22.0 tok/s | |
| 5 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB | 56.1 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | GPT OSS 20BOpenAIHow to run → Q4_K_M · 13.1 GB | 24.4 tok/s | |
| 2 | Q4_K_M · 2.3 GB | 87.4 tok/s | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 22.6 tok/s | |
| 4 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB | 22.0 tok/s | |
| 5 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB | 56.1 tok/s |
| # | Model | Speed ↓ | Score |
|---|---|---|---|
| 1 | GPT OSS 20BOpenAIHow to run → Q4_K_M · 13.1 GB | 24.4 tok/s | |
| 2 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB | 56.1 tok/s | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB | 22.6 tok/s | |
| 4 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB | 22.0 tok/s | |
| 5 | Gemma 4 E2B ITGoogle Q4_K_M · 3.5 GB | 20.9 tok/s |
| # | Model | Speed | Score |
|---|---|---|---|
| 1 | Q4_K_M · 16.5 GB | 22.6 tok/s | |
| 2 | Gemma 3 4B ITGoogle Q4_K_M · 5.0 GB | 25.2 tok/s | |
| 3 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB | 11.3 tok/s | |
| 4 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB | 22.0 tok/s | |
| 5 | Cosmos Reason2 2BNVIDIA Q4_K_M · 3.5 GB | 44.5 tok/s |
The MacBook Air 15" M3 is a 24 GB laptop with more memory than speed: it runs 27–31B models at 4 tok/s. macOS keeps about 3.5 GB, leaving 20.5 GB for models. Prompt reading is its weak spot: a 4K-token prompt waits 26 s before the first word on Gemma 4 26B A4B IT. M3 generation, released 2024; the newer M4 writes 14% faster with 1.2× the compute.
MacBook Air 15" M3 strengths and weaknesses
Strengths
- Runs 35B MoE models: Qwen3.6 35B A3B at 22.6 tok/s in 19.8 GB
Weaknesses
- 70B models don't fit: Llama 3.3 70B Instruct needs 51.3 GB at Q4 vs 20.5 GB usable here
- Slow on large dense models: Qwen3.8 27B at 4.1 tok/s
- Slow prompt reading: A 32K-token prompt on Gemma 4 26B A4B IT waits 3.4 min for the first token
- No CUDA: vLLM, AWQ, GPTQ and FP8 builds don't run; use GGUF or MLX
Memory for models
20.5 GB of 24 GB70B models need 51.3 GB at Q4, more than the 20.5 GB available.
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 Qwen3.6 35B A3B.
Output speed depends on memory bandwidth (102 GB/s). Grey rows need more than the 20.5 GB available.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB11.3tok/s
- 7.5tok/s
- 4.1tok/s
- 22.6tok/s
- 70BLlama 3.3 70B Instruct · does not fitneeds 51.3 GB
- 120B MoEGPT OSS 120B · does not fitneeds 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: 24.9 to 31.1 tok/s since the M1. Compute, which sets how fast prompts are read, grew 3.1×. Indexed to the M1; speeds on Qwen3.6 35B A3B, same 16 GB.
All 1068 models on the MacBook Air 15" M3
Featured onlyBest quant that fits in 20.5 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 14B InstructChatCodeFunctions | Q4_K_M | 12.4 GB | 7.5 t/s | 33K | B45, Decent | |
Llamatron 8B V1Chat | Q4_K_M | 7.3 GB | 13.3 t/s | 131K | B53, Decent | |
| BF16 | 1.2 GB | 68.9 t/s | 131K | S86, Runs great | ||
| BF16 | 2.5 GB | 55.1 t/s | — | A82, Runs well | ||
Mamba 130M HFChat | Q4_K_M | 0.9 GB | 796.4 t/s | — | S100, Runs great | |
XiaoHong V1Chat | Q4_K_M | 7.6 GB | 13.0 t/s | 41K | B52, Decent | |
MobileLLM R1.5 950MChatReasoning | BF16 | 3.5 GB | 34.3 t/s | — | A73, Runs well | |
InternVL2 2BVision | BF16 | 6.6 GB | 14.8 t/s | — | B55, Decent | |
| Q4_K_M | 2.2 GB | 101.0 t/s | — | S92, Runs great | ||
Qehwa Pashto LlmChat | Q4_K_M | 5.8 GB | 14.4 t/s | 131K | B54, Decent | |
Schematron 3BChat | Q4_K_M | 4.1 GB | 29.1 t/s | 131K | A70, Runs well | |
| Q4_K_M | 3.6 GB | 41.4 t/s | — | A76, Runs well | ||
| Q4_K_M | 3.4 GB | 45.1 t/s | 33K | A77, Runs well | ||
| Q4_K_M | 11.8 GB | 7.5 t/s | 41K | B45, Decent | ||
Medgemma 1.5 4B ITVision | Q4_K_M | 5 GB | 25.2 t/s | — | A66, Runs well |
Specs
- Chip
- M3
- CPU cores
- 8
- GPU cores
- 10
- Unified memory
- 24 GB
- For models
- 20.5 GB (macOS keeps 3.5 GB)
- Memory bandwidth
- 102.4 GB/s
- FP16 (dense)
- 7 TFLOPS
- Neural Engine
- 18 TOPS
- Released
- 2024-03-08
MacBook Air 15" M3 FAQ
How much memory is available for AI on the MacBook Air 15" M3 (24 GB)?
About 20.5 GB of the 24 GB. macOS keeps roughly 3.5 GB for itself and other apps.
What is the largest model the MacBook Air 15" M3 (24 GB) can run?
Qwen3.6 35B A3B (36.0B) at IQ4_XS uses 19.8 GB, 97% of the available memory, and runs at about 22.6 tok/s.
Can the MacBook Air 15" M3 (24 GB) run Gemma 4 26B A4B IT?
Yes. Gemma 4 26B A4B IT fits at Q4_K_M, using 16.5 GB of the 20.5 GB available, and it is the best model that runs well on this machine.
Can the MacBook Air 15" M3 (24 GB) run 70B models?
No. Llama 3.3 70B Instruct needs 51.3 GB at Q4, more than the 20.5 GB available.
How long does the MacBook Air 15" M3 (24 GB) take to read a long prompt?
On Gemma 4 26B A4B IT, about 26 s for a 4K-token prompt and 3.4 min for a 32K-token document before the first word appears.
Is renting a GPU cheaper than buying the MacBook Air 15" M3?
An RTX A5000 (24 GB) rents from $0.16/hr.