Mac configuration
Previous generationMacBook Air M3 · 16 GB
16 GB unified memory · 100 GB/s memory bandwidth · verified
A fanless 16 GB Mac. Short local runs are practical; sustained inference can be limited by thermal behavior as well as memory.
What runs comfortably at 8K
| Model | Quantization | Requirement | Context ceiling | MLX command | Ollama command |
|---|---|---|---|---|---|
| Flint 40B39.8B total / ~2.5B active | UD-IQ1_M | 13.1 GB | 109,776 tokens | MLX package not verified | No verified tag |
| NVIDIA Nemotron 3.5 Bolt E4B8B total / 4B effective | Q4_K_M | 8.0 GB | 65,536 tokens | MLX package not verified | ollama run nemotron-3.5-bolt:e4b |
| LFM3-1.7B1.73B dense | Q8_0 | 3.4 GB | 262,144 tokens | mlx_lm.generate --model LiquidAI/LFM3-1.7B-MLX-4bit --prompt "Your prompt" --max-tokens 8192 | ollama run lfm3:1.7b-q8_0 |
| Bonsai 227.36B dense / ternary | PQ2_0 | 9.2 GB | 111,288 tokens | mlx_lm.generate --model prism-ml/Ternary-Bonsai-2-27B-mlx-2bit --prompt "Your prompt" --max-tokens 8192 | No verified tag |
| Gemma 4 12B Unified11.95B dense | Q4_K_M | 9.2 GB | 146,528 tokens | mlx_lm.generate --model google/gemma-4-12B-it --prompt "Your prompt" --max-tokens 8192 | ollama run gemma4:12b-it-q4_K_M |
| Gemma 4 E4B4.5B effective / 8B stored | Q8_0 | 11.0 GB | 131,072 tokens | mlx_lm.generate --model google/gemma-4-E4B-it --prompt "Your prompt" --max-tokens 8192 | ollama run gemma4:e4b-it-q8_0 |
| Qwen3 8B8.2B dense | Q8_0 | 11.6 GB | 32,768 tokens | MLX package not verified | ollama run qwen3:8b-q8_0 |
| DeepSeek R1 8B8B dense | Q8_0 | 11.1 GB | 45,778 tokens | MLX package not verified | ollama run deepseek-r1:8b-llama-distill-q8_0 |
| Llama 3.2 3B3.2B dense | Q8_0 | 5.8 GB | 96,786 tokens | MLX package not verified | ollama run llama3.2:3b-instruct-q8_0 |
| Phi-4 Mini3.8B dense | Q8_0 | 6.7 GB | 79,348 tokens | MLX package not verified | ollama run phi4-mini:3.8b-q8_0 |
| Mistral 7B7.2B dense | Q8_0 | 10.3 GB | 32,768 tokens | MLX package not verified | ollama run mistral:7b-instruct-v0.3-q8_0 |
| Gemma 3 4B4B dense | Q8_0 | 7.6 GB | 68,212 tokens | MLX package not verified | ollama run gemma3:4b-it-q8_0 |
Tight fits and limits
These models are estimated to fit, but with less than 20% headroom. They are not the best first purchase target.
- gpt-oss-20b MXFP4 · 13.8 GB · 95,740 tokens ceiling
Quantization trade-offs for the five featured models
The same 8K formula is applied to each published quantization. A lower-bit file can fit while a higher-quality file does not; that is a capacity result, not a promise that every quantization has the same output quality or backend support.
| Model | Quantization | 8K requirement | Fit | Context ceiling | Ollama / LM Studio |
|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | 13.8 GB | Tight fit | 95,740 tokens | ollama run gpt-oss:20blms get "gpt-oss-20b"@mxfp4 |
| Qwen3.6 35B-A3B | MXFP4 | 24.3 GB | Won't fit | 0 tokens | ollama run qwen3.6:35b-a3blms get "Qwen3.6 35B-A3B"@mxfp4 |
| Qwen3.6 35B-A3B | Q4_K_M | 24.7 GB | Won't fit | 0 tokens | ollama run qwen3.6:35b-a3b-q4_K_Mlms get "Qwen3.6 35B-A3B"@q4_k_m |
| Qwen3.6 35B-A3B | Q8_0 | 39.5 GB | Won't fit | 0 tokens | ollama run qwen3.6:35b-a3b-q8_0lms get "Qwen3.6 35B-A3B"@q8_0 |
| Qwen3.8 27B | Q4_K_M | 19.4 GB | Won't fit | 0 tokens | ollama run qwen3.8:27b-q4_K_Mlms get "Qwen3.8 27B"@q4_k_m |
| Qwen3.8 27B | Q8_0 | 32.0 GB | Won't fit | 0 tokens | ollama run qwen3.8:27b-q8_0lms get "Qwen3.8 27B"@q8_0 |
| Gemma 4 12B Unified | Q4_0 | 9.1 GB | Comfortable fit | 149,498 tokens | ollama run gemma4:12b-it-qatlms get "Gemma 4 12B Unified"@q4_0 |
| Gemma 4 12B Unified | Q4_K_M | 9.2 GB | Comfortable fit | 146,528 tokens | ollama run gemma4:12b-it-q4_K_Mlms get "Gemma 4 12B Unified"@q4_k_m |
| Gemma 4 12B Unified | Q8_0 | 14.7 GB | Tight fit | 33,667 tokens | ollama run gemma4:12b-it-q8_0lms get "Gemma 4 12B Unified"@q8_0 |
| Gemma 4 E4B | Q4_K_M | 7.7 GB | Comfortable fit | 131,072 tokens | ollama run gemma4:e4b-it-q4_K_Mlms get "Gemma 4 E4B"@q4_k_m |
| Gemma 4 E4B | Q4_0 | 7.9 GB | Comfortable fit | 131,072 tokens | ollama run gemma4:e4b-it-qatlms get "Gemma 4 E4B"@q4_0 |
| Gemma 4 E4B | Q8_0 | 11.0 GB | Comfortable fit | 131,072 tokens | ollama run gemma4:e4b-it-q8_0lms get "Gemma 4 E4B"@q8_0 |
Why the unified-memory number is not all yours
The calculator compares the model requirement with the configuration's unified memory, but macOS and the GPU driver reserve memory before the model starts. Closing other applications and reducing the context is more important on a 16 GB Mac than on a dedicated 24 GB GPU with its own VRAM.
The speed band is relative, not a measured tokens-per-second result. The model's active parameter count and the chip's 100 GB/s memory bandwidth are useful for comparison, but backend kernels and thermals still change the outcome.
Low-cost setup upgrades
Accessories that make MacBook Air M3 · 16 GB easier to live with
These optional upgrades improve storage, connectivity, or power resilience. None increases unified memory or turns an oversized model into a fit.
Mac mini dock or stand with an NVMe slot
Adds storage and front-facing ports without buying a larger internal drive. Verify the Mac model, Thunderbolt version, and power-delivery limits before ordering.
Search current listings on AmazonExternal NVMe SSD for model storage
Keeps large GGUF and MLX files off the boot drive. Fast storage reduces load time, but it does not increase the memory available for inference.
Search current listings on AmazonThunderbolt or USB4 NVMe enclosure
A reusable enclosure lets you choose the drive capacity separately. Confirm the host supports the advertised Thunderbolt or USB4 generation.
Search current listings on AmazonAs an Amazon Associate I earn from qualifying purchases. Links are sponsored search links and do not change the price you pay. Affiliate disclosure.
If the model does not fit
Start by reducing context and choosing the lowest published quantization. The next choices are a smaller model or a Mac configuration with more unified memory. External SSDs are useful for storing downloads, but they do not add inference memory.
Upgrade target
More memory, or more bandwidth?
If your models fit but generation feels slow, prioritize a higher-bandwidth chip tier. If the required model does not fit, prioritize a configuration with more unified memory. Do not treat either number as a live price quote.
Search current listings on AmazonAs an Amazon Associate I earn from qualifying purchases. This is a search link, not a live price or availability guarantee. Affiliate disclosure.
Also consider
FAQ
Mac questions, answered.
How much unified memory can I use for a local model on MacBook Air M3 · 16 GB?
This profile uses 16 GB of unified memory as the model budget. macOS, the display, browser tabs, and other applications share the same pool, so the actual runtime headroom is lower.
Does memory bandwidth matter on MacBook Air M3 · 16 GB?
Yes. Apple lists 100 GB/s for this chip tier. Memory capacity decides whether a model can fit; memory bandwidth and GPU compute influence how quickly it generates after it fits.
Why do some pages say MLX is unavailable?
MLX requires a compatible model package and a supported runtime path. If the model record does not list MLX, this guide does not invent a command; use Ollama or llama.cpp after checking the linked source.
Does lower-bit quantization make MacBook Air M3 · 16 GB faster?
Lower-bit weights reduce capacity pressure and can reduce memory traffic, but this page labels speed as an estimate rather than a benchmark. On MacBook Air M3 · 16 GB, 100 GB/s memory bandwidth and backend kernel support still matter more than the quantization label alone.
What should I do if a model still does not fit on MacBook Air M3 · 16 GB?
Reduce context first, then choose a smaller model or a configuration with more unified memory. External storage can hold model files, but it cannot add unified memory. Configurations that fit this machine include Bonsai 2, gpt-oss-20b, Gemma 4 12B Unified, Qwen3 8B.