RAM vs VRAM for Local LLMs
Apple unified memory and GPU VRAM hold the same weights and context cache, but with different budget rules and very different speed per gigabyte. One page, both sides, computed by the same engine the calculator runs.
VRAM and Apple unified memory play the same role — they hold the model's weights plus the KV context cache — under different budget rules and at different speeds. A discrete card hands about 90% of its VRAM to the model, while macOS keeps a share of unified memory, so ModelFit budgets 70–85% of a Mac depending on the tier. Per gigabyte, a GPU streams weights far faster: on the same shared model the engine estimates ~118 vs ~22 tokens/sec between an RTX 5090 and a 32 GB Mac. The trade-off is scale: a Mac's memory ceiling runs to 128 GB and beyond, so high-RAM Macs fit bigger models than any consumer card, while a GPU is faster at the capacity it has. Everything below is derived at build time from ModelFit's 106-model catalog.
With 32 GB of unified memory, ModelFit budgets about 22 GB for the model and comfortably runs local LLMs up to ~35B parameters at Q4. The best single pick is Gemma 4 26B-A4B.
Tokens/sec are ModelFit estimates from chip bandwidth and model size, not measured benchmarks. Ollama commands are registry-verified.
Same models, two memory kinds
A local model's memory bill is the same on both sides: weights at Q4 plus the KV cache for your context. What differs is the pool. Each row pairs the capacity classes both sides actually offer and shows the largest local model that fits under each side's budget rule — the tiered 70–85% share of unified memory on the Apple side, 90% of VRAM on the GPU side.
| Memory class | Apple unified memory | Discrete GPU (VRAM) |
|---|---|---|
| 8 GB | ~5.6 GB usable | ~7.2 GB usable · RTX 4060 |
| 16 GB | ~11.2 GB usable | ~14.4 GB usable · RTX 5080 LFM2 24B-A2B Instruct 24B, 2B active |
| 24 GB | ~16.8 GB usable Granite 4.2 30B 29.3B | ~21.6 GB usable · RTX 4090 Qwen3.5 35B-A3B Instruct 35B, 3B active |
| 32 GB | ~22.4 GB usable Qwen3.5 35B-A3B Instruct 35B, 3B active | ~28.8 GB usable · RTX 5090 Qwen3.6 35B-A3B 35B, 3B active |
| 64 GB | ~48 GB usable | — |
| 96/128 GB | ~76.8 GB usable Qwen3.5 122B-A10B Instruct 122B, 10B active | ~86.4 GB usable · RTX PRO 6000 Qwen3.5 122B-A10B Instruct 122B, 10B active |
Apple side: largest model that fits ModelFit's tiered budget (70% up to 32 GB, rising to 85% at 128 GB and above) — same predicate as the stats tier table. GPU side: largest engine-ranked local model on the representative card of each VRAM class (the same cards the calculator snaps to), at the 90% budget. A MoE is annotated with its active parameters — its total is not comparable to a dense model's. 96 GB and 128 GB Macs land on the same largest model; Mac Studios go up to 512 GB.
Computed from ModelFit's 106-model catalog. Fit and tok/s are ModelFit estimates from memory bandwidth and model size, not measured benchmarks. Updated 2026-09-18.
Speed per GB is not equal
Tokens come out as fast as the memory system can stream the model's weights, which is why the same amount of memory is not the same speed. A card's VRAM is built for that streaming — the RTX 5090 moves 1,792 GB/s, a figure our GPU catalog tracks directly. A Mac's unified memory serves the CPU, the GPU and every open app from one pool, so its effective bandwidth per usable gigabyte is far lower. Take one model both sides can run at 32 GB — the engine's top shared pick:
| Same model, 32 GB on both sides | Estimated speed |
|---|---|
| Gemma 4 26B-A4B 26B, 4B active | — |
| Mac Mini M6 32 GB (unified memory) | est. 22 tok/s |
| RTX 5090 32 GB (VRAM) | est. 118 tok/s |
Same capacity class, same model, roughly 5x the tokens per second: that is bandwidth, not capacity. If your workload is a model that fits a card, a GPU reads its weights faster. If your goal is the biggest model that fits your budget, the Mac's larger ceiling wins instead — that trade is what the table above and the GPU hub vs the RAM guide lay out side by side.
Which should you buy?
If you already own a Mac, buy nothing: your unified memory is your LLM memory, and the tier budget above plus the calculator below size a pick to exactly what you have. If you are buying new and local LLM speed per gigabyte is the goal, a graphics card wins at every class both sides offer. If the biggest dense models (70B class and up) are the goal, a high-RAM Mac or a pooled multi-GPU rig is the only path — no single consumer card holds a dense 70B at Q4, while a 64 GB Mac fits one comfortably. The calculator's multi-GPU mode shows what pooling cards buys you, and Mac memory cannot be added after purchase, so size that tier before you pay for it.
Frequently asked questions
Is unified memory the same as VRAM?
No, but for local LLMs they do the same job: both hold the model's weights and the KV context cache, sized by the same per-parameter math. The difference is the pool. A card's VRAM is dedicated to the GPU, so about 90% of it can go to the model. A Mac's unified memory is one bank shared with macOS and your apps, so the usable share is the tiered 70%-85% budget — a 32 GB Mac has roughly 22.4 GB of model budget against about 28.8 GB on a 32 GB card.
Can a Mac run models that no consumer GPU can?
Yes, at high RAM configurations. In the table above, a 64 GB Mac fits Llama 3.1 70B Instruct (70B), and a 96/128 GB Mac fits Qwen3.5 122B-A10B Instruct (122B total). Consumer cards stop at 32 GB, where the largest dense fit on an RTX 5090 is Ornith 1.0 35B (35B). Only the 96 GB workstation card (RTX PRO 6000) reaches the same territory on the GPU side, and Mac Studios go further still, up to 512 GB.
Why is a GPU faster than a Mac with the same memory?
Decode speed is memory-bound: every token streams the model's weights through memory, so bandwidth decides how fast tokens come out. On the same Gemma 4 26B-A4B (26B total, 4B active), the engine estimates about 22 tokens/sec on a 32 GB Mac Mini M6 against 118 tokens/sec on a 32 GB RTX 5090 — roughly 5x faster. A card's VRAM is built for that streaming (1792 GB/s on the 5090), while a Mac's unified memory serves the CPU, the GPU and every open app from one pool, so less of it is moving model weights at any moment.
Should I buy more Mac RAM or a graphics card?
If you already own a Mac, buy nothing: the unified memory you have is your LLM memory, and the tier budget above plus the calculator below size a pick to it. If you are buying new and want the fastest tokens per gigabyte for the models both sides can run, a discrete GPU wins — more dedicated bandwidth, with a 90% budget against the 70-85% tiered share on a Mac. If you need the biggest dense models (70B class and up), you need a high-RAM Mac or pooled cards, because no single consumer GPU holds a dense 70B at Q4. Decide by which of those two goals matters more, and remember Mac memory is fixed at purchase — it cannot be added later.
Go deeper
The GPU side in depth: per-card picks, usable budgets, and what long context costs.
Weights, KV cache, and the usable budget arithmetic for either kind of memory.
Unified memory, single card and multi-GPU rigs in one tool — every mode here.
The model-size-to-memory matrix on the Apple side, tier by tier.
What fits on your card at each VRAM amount, engine-derived.
The GPU memory bill for popular model sizes, 7B through 70B.
More surfaces: devices hub · GPU hub · can I run it · quantization comparison · open dataset · hardware stats · model pages · guides · how much VRAM for LLMs · best LLM for a 16 GB Mac · Mac Studio · models for a 32 GB tier
The weekly local-AI refresh
New open-weight models, real Apple Silicon benchmarks, and the one model worth running on your Mac this week. Free, one email a week, unsubscribe anytime.
By subscribing you agree to our Privacy Policy and to receive the weekly email. Unsubscribe anytime.