Home / GPUs

Best local LLMs for your GPU

Pick your card. Each guide lists the models that actually fit, with the exact quant — and for 22 of these 24 cards, median tokens/sec measured by real TurboLLM users on that exact GPU, sample size shown on every number. Not sure what you have? Try the VRAM fit checker.

8 GB Small dense models at a mid quant, or a Mixture-of-Experts model with its experts offloaded to system RAM.
12 GB Comfortable for 9–14B dense at Q4–Q5, and MoE models with a moderate expert offload.
16 GB The current sweet spot — a 27B dense model fits at a mid quant, and a 35B-A3B MoE runs fast.
24 GB Dense models up to ~31B fully in VRAM; big MoE models with expert offload.
32 GB and up Frontier-class local models, long contexts, and unified-memory Apple Silicon.

Where these numbers come from

TurboLLM users who opt into anonymous benchmark sharing contribute the measurements behind these pages — model, quant, hardware and measured speed, nothing else. The data refreshes weekly, so coverage grows as more people run more models. Synthetic auto-tune sweeps are kept separate from measurements taken during real use, and every figure is published with its sample size, however small.

$ npx turbollm

One command detects your GPU, provisions a matching engine, and benchmarks each model on your card as it loads. See Install & first run, the model hub, or Quantization explained.