What fits my GPU?
Short answer: Pick your hardware once and see every model in the dataset that fits, with the best quantization for each, the memory left over and the speed. It uses the same math as the single-model checker. However, most speeds are still missing or estimated, so treat memory fit as the firm answer and speed as the soft one.
Data checked 2026-10-22 · memory figures are estimates; measured speeds link their source
Pick your hardware once and see every model in our dataset that fits, with the best quant for each, how much memory is left, and speed (measured where we have it). Same math as the single-model checker.
📊 Most speeds here are missing or estimated. Run llama-bench and paste the output, and it goes in after review.
Loading…
Get the weekly "new models & what it takes to run them" email
Free. One email a week. Unsubscribe anytime.
FAQ
What LLM can I run on my GPU?
Pick your GPU, Mac or Strix Halo here and the list shows every model that fits fully or needs offload, with the best quant for each. For a quick answer by memory size, see the Best LLM for your VRAM pages from 8 to 80 GB.
Does context length change what fits?
Yes. The KV cache grows with context, so a model that fits at 8k tokens can fail at 128k. You can set context from 4k to 256k tokens and pick an f16, q8_0 or q4_0 KV cache.
What happens if a model doesn't fit in VRAM?
It can still run with offload to system RAM, slowly. "Runs slowly with offload" means the model fits in VRAM plus system RAM minus ~6 GB for the OS.
How are mixture-of-experts (MoE) models handled?
We follow llama.cpp's --cpu-moe placement: attention, shared experts, embeddings and the KV cache stay on the GPU, routed experts go to system RAM, and spare VRAM takes back some experts.