On a 192 GB Mac, 12 catalog models run fully on the GPU at an 8,192-token context. The most capable is GLM-4.7 at UD-Q2_K_XL (needs ~142.2 GiB). Pick a smaller model or a lower quant for more headroom.
Figures assume a 192 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 192 GB Mac runs the largest open-weight mixture-of-experts models in our catalog that fit on one machine at all — the biggest at low-bit quants, the class below them at healthier ones. This is effectively an MoE tier: models this size only stay responsive because a fraction of their parameters activate per token. 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 192 GB machine budgets against roughly 144. The list below is derived live from the fit engine, largest first, and it moves as the catalog does.
Three honest uses. First, reach: models that otherwise mean multi-GPU rigs can load here in one machine, at low-bit quants — treat that as the rule-of-thumb tradeoff it is. Second, quant upgrades: the very large MoE class runs at healthier quants than the 128 GB tier allows, which is often the better-quality outcome in practice. Third, context: the KV cache of a giant model is enormous, and this is the first tier where long-context sessions on very large models have real room.
The pool is still shared with macOS, and the GPU's default share of it is still smaller than the machine's total. Even here, the biggest entries want a quiet machine.
GLM-4.7 at UD-Q2_K_XL · 358.3B params · needs ~142.2 GiB
One command to run it (llama.cpp):
llama-server -m GLM-4.7-UD-Q2_K_XL-00001-of-00003.gguf -c 8192 -ngl 999| Model | Sweet-spot quant | Fits in |
|---|---|---|
| GLM-4.7 zai-org | UD-Q2_K_XL Runs fully on GPU | ~142.2 GiB |
| Qwen3-235B-A22B-Thinking-2507 Qwen | UD-Q4_K_XL Runs fully on GPU | ~139.8 GiB |
| 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 |
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.
Some of the largest MoE models we track move from out of reach to fitting at low-bit quants, and the class below them steps up from aggressive low-bit quants to healthier ones. The quieter gain is context headroom: giant models have giant KV caches, and this is the tier where long-context work on them stops being theoretical.
Often, but treat it as a tradeoff, not a free lunch. As a rule of thumb, very large models tolerate aggressive quantization more gracefully than small ones — yet a model one class down at a healthy quant can still win on your task. The list shows both options fit here; test them head to head before committing.
For the mixture-of-experts entries, generally yes — only a fraction of the parameters activate per token, so per-token memory traffic stays far below the total size. Speed still tracks your chip's memory bandwidth, and this page ranks by fit, not speed, so check your specific chip's page before setting expectations.
More than ever — the KV cache scales with both context and model size, and on flagship-scale models it grows fast enough to erase what looks like a huge margin. The verdicts hold at the stated context; if long context is why you bought this machine, verify the exact model and context in the fit calculator.
Not all of it. The pool is shared with macOS and your apps, and macOS caps how much of it the GPU may claim by default. Advanced users can raise that limit for the biggest entries, but leave the OS a working slice — a model that starves the system takes the whole machine down with it.
selected to compare · pick at least 2