What can I do with my GPU?
RigFact shows the captured evidence it holds for the GPU and goal you select — what was tested, what actually happened, and what is still unknown. Pick the GPU you own and RigFact shows the candidate model configurations it can currently reason about, with what is KNOWN, what is still UNKNOWN, and whether a real run command exists. Each configuration's memory requirement or estimate is compared against your GPU's memory. A candidate is not a universal recommendation — every recipe carries its own evidence state.
Scope: Agentic coding. RigFact currently evaluates these recipes for local agentic-coding use; this scope does not yet rank or filter recipes.
What these words mean
• Quantization — a way to make a model smaller so it uses less memory; different methods may affect quality and performance.
• Context — how much information the model can work with at one time.
• Tokens — small pieces of text used by the model.
• KV cache — memory used while the model keeps track of its active context.
• Recipe — the exact model + settings + runtime configuration RigFact tested.
• Reproducible — enough technical identity was captured to seriously attempt the same run again; it is not a guarantee.
Technical terms like these appear inside each card's Technical details.
How does this work?
• Each configuration is checked on its own.
• Memory is shown in decimal GB, the same units as your GPU's published memory.
• "Fits the known GPU memory requirement" means the memory need is met — not that the model will run.
• "Not enough information" means a required fact about your system could not be resolved.
• Recipes state what RigFact actually knows. Unknown facts stay unknown — they are never guessed.
• "REPRODUCIBLE" on a recipe means it carries enough exact identity, source, and execution facts to attempt a serious reproduction of that run. It is not an endorsement, not a promise of success, and not a claim of universal fit — each recipe stands on its own evidence.