Three-tier GPU matching
PCI device identity first, vendor tools second, normalized fuzzy matching last—with confidence carried through the result.
← All workOPEN SOURCE / SYSTEMS TOOLING
CIRI — Can I Run It?
Which open-source LLMs can this machine run—and how fast? CIRI answers with detected hardware, honest memory math, and benchmark-aware estimates.
GO · BUBBLE TEA · GHW · EMBEDDED DATA5,348model catalog entries
149GPU profiles
3prediction tiers
MITopen-source license
THE PROBLEM
Model compatibility should not be a guessing game.
Model pages list parameters and quantization, but users still need to translate that into VRAM pressure, system-memory spill, and expected tokens per second on their exact hardware.
CIRI brings hardware detection, a model catalog, benchmark data, and a transparent prediction cascade into one cross-platform terminal interface.
SYSTEM ARCHITECTURE
Detection, data, prediction, interface.
Honest memory assessment
Fit uses the larger of catalog VRAM and calculated weight size, adds a 10% buffer, and treats Apple unified memory separately.
Evidence before heuristic
Speed estimates prefer exact benchmark matches, then architecture-family scaling, then a transparent bandwidth/compute heuristic.
OPEN SOURCE
Inspect the implementation.
CIRI is MIT licensed. Its datasets are embedded into the binary, so the core workflow runs without a hosted service.
go install github.com/cezaryt5/ciri/cmd/ciri@v1.0.0