On a 16 GB Mac, 12 catalog models run fully on the GPU at an 8,192-token context. The most capable is Kimi-VL-A3B-Instruct at IQ4_XS (needs ~9.7 GiB). Pick a smaller model or a lower quant for more headroom.
Figures assume a 16 GB unified-memory Mac (one pool shared by CPU and GPU), judged at an 8,192-token context. Speeds depend on the specific chip — see the GPU pages.
TL;DR
A 16 GB Mac runs the 7B and 8B instruct and reasoning class fully on the GPU at 4-bit sweet-spot quants, with enough slack for real context. The 13B–14B class is the edge rather than the comfortable option: macOS hands the GPU only part of the unified pool by default — roughly three quarters, and about two thirds on machines at 32 GB and below, so a 16 GB machine is budgeting against something nearer 10. The list below is derived live from the fit engine, largest first.
On a discrete GPU, a model that barely fits still has the card to itself. On a Mac it doesn't — the pool is shared with macOS and your apps, and the GPU's default share of it is smaller than the machine's total. So a top-of-list entry here is a "close everything and load it" experience rather than a daily driver.
The dependable move at this tier is one notch down: a 7B or 8B model at its sweet-spot quant, with room for long context and a normal desktop. Save the biggest entry for when you actually need it.
Kimi-VL-A3B-Instruct at IQ4_XS · 16.4B params · needs ~9.7 GiB
One command to run it (llama.cpp):
llama-server -m Kimi-VL-A3B-Instruct.i1-IQ4_XS.gguf -c 8192 -ngl 999| Model | Sweet-spot quant | Fits in |
|---|---|---|
| Kimi-VL-A3B-Instruct moonshotai | IQ4_XS Runs fully on GPU | ~9.7 GiB |
| DeepSeek-R1-Distill-Qwen-14B deepseek-ai | IQ4_XS Runs fully on GPU | ~10.3 GiB |
| Qwen3-14B Qwen | IQ4_XS Runs fully on GPU | ~10.1 GiB |
| Phi-4-reasoning microsoft | IQ4_XS Runs fully on GPU | ~10.2 GiB |
| phi-4 microsoft | IQ4_XS Runs fully on GPU | ~10.2 GiB |
| gemma-4-12B-it google | Q4_K_M Runs fully on GPU | ~8.6 GiB |
| Qwen3-8B Qwen | Q8_0 Runs fully on GPU | ~10.6 GiB |
| DeepSeek-R1-Distill-Llama-8B deepseek-ai | Q8_0 Runs fully on GPU | ~10.3 GiB |
| gemma-4-E4B-it google | Q8_0 Runs fully on GPU | ~9.1 GiB |
| DeepSeek-R1-Distill-Qwen-7B deepseek-ai | Q8_0 Runs fully on GPU | ~9.3 GiB |
| Qwen3-Embedding-8B Qwen | Q4_K_M Runs fully on GPU | ~6.4 GiB |
| gemma-4-E2B-it google | Q8_0 Runs fully on GPU | ~5.8 GiB |
Derived live from the fit engine + catalog at an 8,192-token context. "Fits in" is the modelled VRAM the sweet-spot quant needs (weights + KV cache + overhead). Speed depends on your specific card — check a GPU page or the calculator.
The 7B and 8B instruct and reasoning class — the models most local-LLM communities treat as the everyday recommendation — runs fully on the GPU at 4-bit sweet-spot quants here, with room for real context. Smaller models step up to high, near-lossless quants. The 13B–14B class comes into range but sits near the top of what the GPU is allowed to claim, so treat it as the stretch rather than the default.
Realistically, no. The largest entries budget nearly the whole pool, which is shared with macOS and your open apps — and macOS caps the GPU's share of unified memory by default. Load top-of-list models on a quiet machine, and pick something further down the list as your daily driver.
No, not at a quant worth running. That class wants more memory than the GPU is given here once the KV cache and overhead are counted, and crushing one to very low bits usually gives worse output than a well-quantized 7B or 8B model. If the 30B class is the goal, the 36 GB tier is the first place it fits fully on the GPU.
Yes — the KV cache grows with context and shares the pool with the weights, macOS, and your apps. A model that fits at this page's stated context can push the system into swapping at much longer ones, so verify your model and target context in the fit calculator before a big download.
This page ranks by fit, not speed. A 16 GB configuration exists across several Apple Silicon tiers and generations, and generation speed mostly tracks the chip's memory bandwidth — so identical memory can mean quite different speeds. Check your specific chip's page for speed context.
selected to compare · pick at least 2