Apple · 2026 · MacBook Air
Choose the unified-memory configuration, estimate room for your model, then inspect speed and whole-machine price evidence.
Current generation; fanless laptop. Selected published configurations, not an exhaustive Apple configurator. SSD, display, thermal design and GPU-core variants must be matched separately.
12 GiB model/runtime budget after 4 GiB system allowance.
The allowance is editable. Lowering it does not raise macOS’s actual Metal working-set limit. Four sample models, Q4_K_M weights, FP16 KV, one request and 2 GiB runtime reserve. Models above 70B are not assessed here.
Within planning budget
4.8 GiB spare
Qwen/Qwen3-8B-GGUF · revision 7c41481f57cb95916b40956ab2f0b139b296d974. KV architecture ↗
Within planning budget
1.0 GiB spare
Qwen/Qwen3-14B-GGUF · revision 530227a7d994db8eca5ab5ced2fb692b614357fd. KV architecture ↗
Over planning budget
9.4 GiB over budget
Qwen/Qwen3-32B-GGUF · revision 938a7432affaec9157f883a87164e2646ae17555. KV architecture ↗
Over planning budget
30.9 GiB over budget
bartowski/Llama-3.3-70B-Instruct-GGUF · revision b6c5c9f176f3279204034e1d16d393105e95cb88. KV architecture ↗
We do not infer tok/s from memory size, bandwidth, GPU core count or Neural Engine TOPS. The capacity models above have not been benchmarked by us on this Mac.
Changing the memory/context selector changes estimates only. Published references keep their own model, quantization and scope; no exact-configuration speed is implied.
No exact configured purchase quote collected. Confirm the GPU cores, memory, SSD, seller, condition, tax and stock before comparing prices.
Check the official configurations ↗
No matching Mac cloud-compute quote collected. Hardware financing and equipment leases are not GPU-hour rental rates.
Compare the whole-machine budget with rented compute and APIs →
For the GGUF files used in this estimate, llama.cpp provides a Metal backend on Apple Silicon. Download the exact model artifact linked in the capacity cards, set an explicit context, and confirm GPU/Metal execution and actual memory use in the runtime logs. This page does not claim a tested installation on this machine.
MLX is another Apple Silicon runtime path. Its model checkpoints, quantization and memory allocations differ; the GGUF estimates above are not measurements of an MLX checkpoint. Recheck the chosen artifact rather than carrying over a GGUF file size.
llama.cpp build and Metal instructions ↗ · MLX-LM installation and model usage ↗
Apple designs the integrated GPU in this M5 system. Unified memory is shared with the CPU and system; it is not dedicated VRAM. GPU core count and selected memory options come from the linked specification.
Official Apple specifications ↗ · checked 7 Nov 2026.
Hardware and price evidence → · Historical chip benchmark evidence →
These Macs contain an integrated Apple GPU. CPU, GPU, macOS and other apps share one memory pool. We reserve 25% (at least 4 GiB) for the system and other apps, then add exact Q4_K_M weight-file bytes, FP16 KV cache and 2 GiB for the inference runtime. The system allowance is our editable planning assumption, not a measured free-memory value or a universal macOS GPU limit.
Nominal Apple GB labels are treated as GiB for this scenario. Context includes input and output tokens; one request, weights kept in memory. Passing the estimate is a shortlist, not a verified deployment. The four-model sample stops at 70B; a 70B result is not a ceiling on a large Mac.
Actual Metal working-set limits, prefill workspace, app usage, runtime allocations and model support may change the result. Swapping to SSD can make a model load while becoming too slow for the intended task. Memory bandwidth alone cannot establish output speed.
KV bytes = 2 × layers × KV heads × head dimension × context × 2 bytes. 1 GiB = 2³⁰ bytes. Capacity uses the same four pinned GGUF profiles as our GPU catalog.
Qwen3 8B · Q4_K_M: 4.7 GiB; revision 7c41481f57cb95916b40956ab2f0b139b296d974. Architecture ↗
Qwen3 14B · Q4_K_M: 8.4 GiB; revision 530227a7d994db8eca5ab5ced2fb692b614357fd. Architecture ↗
Qwen3 32B · Q4_K_M: 18.4 GiB; revision 938a7432affaec9157f883a87164e2646ae17555. Architecture ↗
Llama 3.3 70B · Q4_K_M: 39.6 GiB; revision b6c5c9f176f3279204034e1d16d393105e95cb88. Architecture ↗
10 GPU cores
Apple · 24/48/64 GB20 GPU cores
Apple · 16/24/32 GB12 GPU cores
All Mac configurations → · Compare discrete graphics cards →