On a 128 GB Mac, 12 catalog models run fully on the GPU at an 8,192-token context. The most capable is Qwen3-235B-A22B-Instruct-2507 at UD-Q2_K_XL (needs ~93.1 GiB). Pick a smaller model or a lower quant for more headroom.
Figures assume a 128 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 128 GB Mac reaches into the very large mixture-of-experts class at aggressive low-bit quants, and runs the class below it at healthier ones. MoE architecture is what makes any of it practical: only a fraction of the parameters activate per token, so total size can be large while per-token work stays manageable. 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 128 GB machine budgets against roughly 96 — which is where the top of this list is decided. The list below is derived live from the fit engine, largest first.
This tier's real decision isn't what fits — it's which tradeoff to take. The largest entries only squeeze in at very low-bit quants, and as a rule of thumb very large models tolerate aggressive quantization better than small ones do, though the cost isn't zero. The alternative is a model one class down at a healthy quant, with the leftover pool spent on long context instead of raw parameter count.
Neither answer is universally right. The list shows what fits; your task decides. Test both before making one your daily driver.
Qwen3-235B-A22B-Instruct-2507 at UD-Q2_K_XL · 235.1B params · needs ~93.1 GiB
One command to run it (llama.cpp):
llama-server -m Qwen3-235B-A22B-Instruct-2507-UD-Q2_K_XL-00001-of-00002.gguf -c 8192 -ngl 999| Model | Sweet-spot quant | Fits in |
|---|---|---|
| Qwen3-235B-A22B-Instruct-2507 Qwen | UD-Q2_K_XL Runs fully on GPU | ~93.1 GiB |
| gpt-oss-120b openai | F16 Runs fully on GPU | ~68.2 GiB |
| Hunyuan-A13B-Instruct tencent | Q8_0 Runs fully on GPU | ~89.2 GiB |
| Kimi-Dev-72B moonshotai | UD-Q5_K_XL Runs fully on GPU | ~58.6 GiB |
| Qwen3.6-35B-A3B Qwen | Q8_0 Runs fully on GPU | ~39 GiB |
| Qwen3-Omni-30B-A3B-Instruct Qwen | Q4_K_M Runs fully on GPU | ~20.3 GiB |
| Qwen2.5-VL-32B-Instruct Qwen | Q8_0 Runs fully on GPU | ~38.3 GiB |
| DeepSeek-R1-Distill-Qwen-32B deepseek-ai | Q8_0 Runs fully on GPU | ~38.3 GiB |
| Qwen3-32B Qwen | Q8_0 Runs fully on GPU | ~38.3 GiB |
| gemma-4-31B-it google | Q8_0 Runs fully on GPU | ~36 GiB |
| GLM-4.7-Flash zai-org | Q8_0 Runs fully on GPU | ~33.6 GiB |
| Qwen3-Coder-30B-A3B-Instruct Qwen | Q8_0 Runs fully on GPU | ~34.6 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.
Not the frontier itself. It reaches into the very large mixture-of-experts class at aggressive low-bit quants, and runs the class below at healthier ones. What it can't do is run the largest models at high-precision quants — that's what the 192 GB tier and multi-machine setups are for. The list below is the current answer, and it changes as the catalog does.
Because weights scale with parameter count, and low-bit quantization is the only way the largest class squeezes into the pool. As a rule of thumb, very large models degrade more gracefully under aggressive quantization than small ones do — but there is still a cost, so compare against a smaller model at a healthier quant on your actual task.
Mixture-of-experts models activate only a fraction of their parameters for each token, so the per-token memory traffic tracks the active share rather than the total size. All the weights still have to fit in the pool — that's what this tier buys — but generation stays usable in a way a dense model of the same total size would not.
Yes, and the effect is largest exactly where you'd hope it wasn't — the KV cache of a very large model grows quickly with context, and the biggest entries here have little slack left for it. The verdicts hold at the stated context; for long-context work on a giant, check the fit calculator first.
This page ranks by fit, not speed. 128 GB configurations sit on the higher Apple Silicon tiers, but bandwidth still varies between them, and it's the main thing separating a comfortable giant from a sluggish one. Check your specific chip's page for speed context.
selected to compare · pick at least 2