Free guide · Aug 24, 2026
Run a 67GB AI model on a 12GB graphics card
Which model fits your machine, the install, the five bugs I hit with fixes, and pointing Claude Code at it.
The short version
FreeToken is an inference engine. It is the thing that *runs* a model, like Ollama or LM Studio. It is not a model itself, and it is not "an uncensored AI." Whether the model refuses you is a property of the model you load, not of FreeToken.
Three things that will save you time:
It only helps with Mixture-of-Experts models. This is the whole trick. A regular ("dense") model has to use every parameter for every word it writes. A Mixture-of-Experts model has a huge total size but only wakes up a small slice for each word. FreeToken is built specifically to exploit that. Point it at a dense model and the magic does not apply.
What's in the full guide
- 012. Why this works at all
- 023. Which model fits your machine
- 034. Installing it
- 045. Every bug I hit, and the fix
- 056. Pick the right quantization format
- 067. Point Claude Code (or Codex) at your local model
- 078. My numbers, and why yours may look nothing like them
- 089. The honest summary
- 09Want the next one of these?