On a 48 GB Mac, 12 catalog models run fully on the GPU at an 8,192-token context. The most capable is Qwen3.6-35B-A3B at Q4_K_M (needs ~23.9 GiB). Pick a smaller model or a lower quant for more headroom.
Figures assume a 48 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 48 GB Mac changes the question from "what fits" to "how well it fits". The 30B class runs at high, near-lossless quants with genuine headroom, and the larger mixture-of-experts models come into range at 4-bit. Dense 70B-class models do not fit here — 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 48 GB machine budgets against roughly 36, and a 70B at 4-bit wants more than that before a single KV token. The list below is derived live from the fit engine, largest first.
The honest way to spend 48 GB is on how well you run a model rather than on how big it is: the 30B class at near-lossless quants, with room for long context and a normal desktop. That is the setup most people end up keeping.
The larger MoE entries are the interesting reach. A dense model streams all of its parameters every token, so even where a large dense model loads it can feel slow on lower-bandwidth chips — as a rule of thumb, an MoE model of similar total size feels quicker. If the dense 70B class is the goal, 64 GB is the first tier where it fits at all.
Qwen3.6-35B-A3B at Q4_K_M · 36B params · needs ~23.9 GiB
One command to run it (llama.cpp):
llama-server -m Qwen3.6-35B-A3B-UD-Q4_K_M.gguf -c 8192 -ngl 999| Model | Sweet-spot quant | Fits in |
|---|---|---|
| Qwen3.6-35B-A3B Qwen | Q4_K_M Runs fully on GPU | ~23.9 GiB |
| Qwen3-Omni-30B-A3B-Instruct Qwen | Q4_K_M Runs fully on GPU | ~20.3 GiB |
| Qwen2.5-VL-32B-Instruct Qwen | Q4_K_M Runs fully on GPU | ~22.9 GiB |
| DeepSeek-R1-Distill-Qwen-32B deepseek-ai | Q4_K_M Runs fully on GPU | ~22.9 GiB |
| Qwen3-32B Qwen | Q4_K_M Runs fully on GPU | ~22.8 GiB |
| gemma-4-31B-it google | Q4_K_M Runs fully on GPU | ~21.3 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 |
| Qwen3-30B-A3B-Instruct-2507 Qwen | Q8_0 Runs fully on GPU | ~34.6 GiB |
| Qwen3.6-27B Qwen | Q8_0 Runs fully on GPU | ~31.9 GiB |
| gemma-4-26B-A4B-it google | Q8_0 Runs fully on GPU | ~28.6 GiB |
| Devstral-Small-2-24B-Instruct-2512 mistralai | Q8_0 Runs fully on GPU | ~27.5 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.
Quality rather than a new class. The 30B class steps up to high, near-lossless quants with real headroom, and the larger mixture-of-experts models come into range at 4-bit. Dense 70B-class models do not fit here — at a 4-bit sweet-spot quant a 70B wants more than the GPU is given on a 48 GB machine, before any KV cache. That starts at 64 GB, and even there it is a tight fit.
As a rule of thumb, parameter count wins on capability — but on a Mac a large dense model costs you twice, in headroom and in speed, since every parameter streams through memory each token. The large mixture-of-experts entries are the middle path: big total size, a fraction of it active per token. Test both on your actual task before committing.
Less than the headline number. The pool is shared with macOS and your open apps, and macOS caps how much of it the GPU may claim by default. Advanced users can raise that limit, but always leave the OS a working slice — the top entries on this page assume a mostly quiet machine either way.
Yes, and more than at smaller tiers — the KV cache grows with context and grows faster on larger models, so the big entries here are the ones most sensitive to it. The verdicts hold at the stated context; for long-document or long-chat work, verify your model and context in the fit calculator first.
This page ranks by fit, not speed. A 48 GB configuration exists across different Apple Silicon tiers, and generation speed mostly tracks memory bandwidth — the difference is most visible on exactly the large dense models this tier newly fits. Check your specific chip's page for speed context.
selected to compare · pick at least 2