How Much RAM, and Which Mac or GPU
The single question behind every local model guide is whether it will fit your machine. This is the reference the other guides point to: the memory math, what a Mac Mini, MacBook Pro, and Mac Studio each hold, the equivalent GPU tiers, and a table mapping models to the hardware that runs them. Read it once and the sizing tables in every model guide make sense.
The memory math
You can estimate what a model needs without downloading it. The weights take roughly the parameter count times bits-per-weight, divided by eight, in bytes. A 4-bit quant is about 4.5 to 5 bits-per-weight, an 8-bit quant about 8.5. So a 32B model at 4-bit is around 18 to 20 GB, a 70B is around 40 to 43 GB, and full 16-bit precision is roughly double the 8-bit figure. On top of the weights you need ten to twenty percent headroom and the KV cache, which grows with how much context you keep. The working rule: your combined VRAM plus system RAM should match or exceed the model file size for six or more tokens per second. It still runs with less, spilling to disk, just slowly.
Apple Silicon: Mini, Pro, Studio
Apple Silicon is unusually good at this because the CPU and GPU share one pool of unified memory. There is no separate VRAM to run out of; the amount of unified memory is the amount you can give a model. That lets a well-specced Mac hold models that no single consumer graphics card could. What differs across the lineup is the ceiling.
| Mac | Memory ceiling | Practical for |
|---|---|---|
| Mac Mini | up to ~64 GB | up to a 70B at 4-bit |
| MacBook Pro | up to ~128 GB | up to a 120B-class model |
| Mac Studio | up to ~512 GB | large MoE and 600B-plus at dynamic quants |
Exact ceilings depend on the chip generation and configuration, so check the current specs, but the shape holds: the Mini is the entry point, the MacBook Pro covers most serious local work, and the Studio is the machine that runs the frontier open models. One caveat: macOS reserves part of unified memory for the system and, by default, caps how much the GPU can claim. On a memory-heavy model you may need to raise that limit; keep total usage under roughly three quarters of your unified memory as a safe target.
What fits which Mac
By unified-memory size, independent of which enclosure it comes in:
| Unified memory | Comfortably runs |
|---|---|
| 16 GB | up to ~8 to 14B at 4-bit; gpt-oss-20b is tight |
| 32 GB | a 32B at 4-bit; gpt-oss-20b easily |
| 48 GB | a 32B at 8-bit, or a 70B at 4-bit (tight) |
| 64 GB | a 70B at 4-bit; gpt-oss-120b; Llama 4 Scout (tight) |
| 128 GB | a 120B-class model; Qwen3-235B and GLM Air at dynamic quants |
| 192 to 256 GB | 235B to 480B mixture-of-experts at dynamic quants |
| 512 GB | DeepSeek 671B and Kimi-class trillion-parameter MoE at dynamic quants |
NVIDIA and PC GPUs
On a PC the number that matters is VRAM, and unlike a Mac you cannot borrow system RAM without a speed penalty, except through the mixture-of-experts offload trick below.
| VRAM | Card | Comfortably runs |
|---|---|---|
| 8 GB | 3060 / 4060 | up to ~8B at 4-bit |
| 16 GB | 4060 Ti / 4080 | a 14B at 4-bit; gpt-oss-20b |
| 24 GB | 3090 / 4090 | a 32B at 4-bit, the consumer sweet spot |
| 48 GB | 2×24GB / A6000 | a 70B at 4-bit |
| 80 GB | H100 / A100 | a 70B at 8-bit; Llama 4 Scout; gpt-oss-120b |
| multi-GPU | server | full-precision 200B-plus MoE and the 600B-plus class |
The mixture-of-experts exception
Everything above assumes a dense model whose weights all sit in fast memory. Mixture-of-experts models change the calculus. They activate only a fraction of their parameters per token, so they compute at the speed of a much smaller model, but they still need memory to hold every expert. That is what makes a trillion-parameter model runnable at home: on a PC, you keep attention on a 24 GB GPU and offload the expert layers to a large pool of ordinary system RAM with llama.cpp's -ot flag. Throughput is then bounded by RAM bandwidth, typically a few tokens per second, but it runs. On a Mac the same models fit directly because unified memory is one big pool. This is why a 24 GB card plus 256 GB of RAM, or a 256 GB Mac, can run a model that a single 80 GB data-center GPU cannot hold.
Model to hardware
| Model | Minimum comfortable |
|---|---|
| Qwen3 32B, Gemma 3 27B | 24 GB VRAM or 32 GB Mac |
| Llama 70B, R1 70B distill | 48 GB VRAM or 64 GB Mac |
| gpt-oss 120B, Llama 4 Scout | 80 GB GPU or 64 to 128 GB Mac |
| Qwen3-235B, GLM 4.6 | 128 GB Mac or multi-GPU, dynamic quant |
| DeepSeek 671B, Kimi K3 | 256 to 512 GB Mac, or 24 GB GPU + 256 GB RAM, or multi-GPU |
How this was checked
The commands in every model guide are the real invocations from each engine's own documentation, not approximations, and I run this stack locally. As a spot check while writing these, on a MacBook Pro with an M1 Max and 64 GB, the Ollama path worked end to end: a one-command pull and run of a small Gemma model downloaded and answered in under a minute. The memory figures above are rules of thumb rounded to be safe; a specific model's exact size and context limit still belong to its model card, which every guide tells you to confirm. Treat the tables as a starting point that gets you to the right ballpark, then verify the one model you are about to run.
FAQ
How much RAM do I need? Roughly parameters times bits-per-weight over eight, plus headroom and KV cache. A 32B at 4-bit is about 18 to 20 GB, a 70B about 40 to 43 GB. Match combined VRAM plus RAM to the file size.
Which Mac is best? By memory ceiling: Mini up to about 64 GB, MacBook Pro up to about 128 GB, Mac Studio up to 512 GB. More unified memory holds larger models.
Can a MacBook Pro run large models? Yes, within its memory. 64 GB runs a 70B and gpt-oss-120b; 128 GB runs a 120B-class model and large MoE at dynamic quants. Trillion-parameter models want a 256 GB-plus Mac Studio or a server.