LLM Hardware Requirements Calculator

Enter your RAM or GPU VRAM and see exactly which local AI models you can run, the single best pick, and how fast it will go.

What LLM can I run? A local model at Q4 quantization needs roughly 0.6 GB of memory per billion parameters. ModelFit budgets about 70% of unified memory for the model on machines up to 32GB, scaling to ~85% at 128GB and above, and about 90% of a discrete GPU's VRAM. That means an 8GB device runs models up to ~9B, a 16GB device comfortably runs up to ~14B, 32GB unlocks ~35B-class models, and 64GB or more runs 70B-class models. Use the calculator below to size your exact hardware against 106 local models.

Requirements calculator

Pick at least two cards (duplicates allowed, e.g. 2x RTX 5060 Ti) to see what the pooled VRAM runs.

Pooling works: Ollama and llama.cpp split layers across cards automatically. With the default layer split, throughput stays near single-card speed and mixed cards run at the slower card's pace. Two identical cards using llama.cpp's row (tensor) split can decode faster than one card, since both memory buses read weights in parallel, though transfer overhead keeps it below 2x. Mixing NVIDIA and AMD in one rig requires the Vulkan build. Fit is what pooling reliably buys you; per-card speed estimates on the GPU pages are single-card only. Long context also eats into the budget (KV-cache), so leave headroom.

Pooled fit only: cross-card tok/s depend on the split mode and the slowest card, so ModelFit does not publish a combined speed estimate. Single-card speeds live on the GPU pages.

Local LLM memory requirements by RAM tier

Every row is derived from ModelFit's catalog of 106 local models across 24 families. Click a tier for the full list.

MemoryModel budget (~70-85%)Max dense model sizeModels that fitTop pick
8 GB~5.6 GB~9B params32 / 106Ornith 1.0 9B
16 GB~11.2 GB~14B params52 / 106Qwen3.5 9B Instruct (Q8)
24 GB~16.8 GB~29.3B params65 / 106Qwen3.8 27B
32 GB~22.4 GB~35B params77 / 106Qwen3.6 35B-A3B
48 GB~34.8 GB~35B params88 / 106Qwen3.8 27B (Q8)
64 GB~48 GB~70B params93 / 106Qwen3.6 35B-A3B (Q8)
96 GB~76.8 GB~70B params99 / 106Qwen3.5 122B-A10B Instruct
128 GB~108.8 GB~70B params101 / 106Qwen3.5 122B-A10B Instruct

Q4_K_M assumed. Fit and tok/s are ModelFit estimates from the dataset, not measured benchmarks. Updated 2026-09-18.

What each memory tier adds

Each card leads with the models that tier unlocks over the tier below, then the best models that already fit, all from the same 106-model catalog as the calculator above. Every model name links to its own memory math.

8 GB

32 OF 106 MODELS FIT

~5.6 GB usable for weights · everything that fits starts here

16 GB

52 OF 106 MODELS FIT

~11.2 GB usable for weights · 5 newly unlocked over 5.6 GB

24 GB

65 OF 106 MODELS FIT

~16.8 GB usable for weights · 5 newly unlocked over 11.2 GB

32 GB

77 OF 106 MODELS FIT

~22.4 GB usable for weights · 5 newly unlocked over 16.8 GB

48 GB

88 OF 106 MODELS FIT

~34.8 GB usable for weights · 5 newly unlocked over 22.4 GB

64 GB

93 OF 106 MODELS FIT

~48 GB usable for weights · 5 newly unlocked over 34.8 GB

96 GB

99 OF 106 MODELS FIT

~76.8 GB usable for weights · 5 newly unlocked over 48 GB

128 GB

101 OF 106 MODELS FIT

~108.8 GB usable for weights · 2 newly unlocked over 76.8 GB

Loads are Q4_K_M estimates including weights only; context and KV-cache need headroom on top. Click a tier for the full ranked list, or a model name for its per-hardware fit math.

Frequently asked questions

How much RAM do I need to run a local LLM?

At Q4 quantization a local LLM needs roughly 0.6 GB of memory per billion parameters, and ModelFit budgets ~70% of unified memory for the model up to 32GB, scaling to ~85% at 128GB and above. In practice 8GB runs models up to ~9B, 16GB comfortably runs up to ~14B, 32GB unlocks ~35B-class models, and 64GB or more runs 70B-class models.

What LLM can I run with my GPU VRAM?

VRAM is the hard ceiling for a discrete GPU, and about 90% of it is usable for model weights. 8GB fits up to an 8B model, 12GB reaches 9B, 16GB unlocks 14B-class models, 24GB runs 32B-class models, and 32GB reaches 70B-class MoE models. Switch the calculator to GPU VRAM mode and enter your card memory to see the exact picks.

How does the calculator work?

Enter your memory amount and pick Apple unified memory or GPU VRAM. The calculator runs ModelFit’s recommendation engine in your browser: it sizes each model at ~0.6 GB per billion parameters, applies the memory budget for your hardware, and ranks the models that fit by quality and speed. A shared link carries its configuration in the URL, and that configuration is rendered on the server first, so /calculator/?gb=32 already shows the 32 GB picks before any JavaScript runs. Tokens per second are ModelFit estimates from memory bandwidth and model size, not measured benchmarks.

Can I combine two GPUs to run bigger local models?

Yes. Ollama and llama.cpp split model layers across cards automatically, so two or three GPUs pool their VRAM for fit: about 90% of the combined VRAM is usable for weights. Expect real throughput below a single card with the same total VRAM, because inter-GPU transfers add overhead and mixed cards run at the slower card’s pace. Switch the calculator to Multi-GPU rig mode to pick your exact cards and see which models fit, and use the Copy link button to share the setup.

Is the ModelFit calculator free?

Yes. The calculator is completely free, needs no sign-up, and runs entirely in your browser with no data sent to a server. The underlying compatibility dataset is open under CC BY 4.0, and the same engine ships as the free npx @wecko-ai/modelfit command-line tool.

Go deeper

PREFER THE TERMINAL?

The same engine runs offline as a one-line command that detects your machine and names the best local model:

npx @wecko-ai/modelfit
LLM Hardware Requirements Calculator (2026) | ModelFit