About

CanItRunLocal answers one question: will a given model actually run on a given machine, and how fast? Two independent tools do that —

• Browse — crowdsourced Hardware × Model × Runtime performance reports.
• Calculator — a deterministic memory-fit estimate you can run for hardware you don't own yet. No account needed.

Every report is submitted by a person who ran the model. Reports are grouped into buckets keyed on a normalized hardware identity × model × runtime, so independent reports on equivalent rigs land together.

A bucket's status:

• Verified — two or more reports whose generation speed agrees within ~10%.
• Conflicting — reports split into more than one cluster; both are shown rather than averaged away.
• Unverified — a single report so far.

• Models — synced daily from Hugging Face GGUF repositories (base model → variant → quant). Never user-submitted.
• Hardware specs (VRAM, memory bandwidth) — synced from a curated local-hardware spec catalog.
• Runtimes — a curated list: llama.cpp, Ollama, LM Studio, vLLM, MLX.

A closed-form estimate: model weights (parameters × bits-per-weight) plus KV cache at the chosen context length, against the rig's usable memory. It never invents data — if a table it needs is empty, it says so rather than guessing. Treat the result as a fit estimate, not a benchmark.

• Reports are self-reported and un-audited; convergence across independent submissions is the only verification.
• The calculator models memory fit, not throughput — a model that fits can still be slow.
• Free-text hardware entries are normalized, not fuzzy-matched, so near-identical rig strings may not yet share a bucket.