About RunLocalAI
RunLocalAI exists to answer one question: can my hardware run this model, and how fast? Most AI coverage is about cloud APIs. This site focuses on the underserved local-inference niche — Ollama, LM Studio, llama.cpp, vLLM, KoboldCPP, MLX, ExLlamaV2 — and the real hardware operators actually own.
Every page is anchored to verifiable data: parameter counts, license terms, quantization sizes, real tokens-per-second measurements. The site does not publish vibes — it publishes numbers measured directly or sourced from named community contributors with reproduction notes.
What this site is for
The local-AI ecosystem changes weekly. Drivers shift, runtimes fork, model architectures land that nobody has benchmarked yet, and yesterday's “best 24 GB card” gets superseded by a used market that didn't exist last quarter. Most operator decisions sit at the intersection of three questions: what hardware do I have? what model do I want to run? what runtime + quant + context combination actually fits?
RunLocalAI is built around answering those three questions honestly. The will-it-run engine gives you the math instantly. The hardware verdicts give you the buyer-grade analysis with skip-warnings + used-market notes. The buyer guides tie those together for the most-asked operator situations.
What this site refuses to do
- Recommend more expensive hardware than the workload actually needs. The buyer guides include “skip this if” sections by design.
- Inflate scores or invent benchmark numbers. Anything we haven't measured ourselves is labeled by provenance: source-backed, community, extrapolated, or estimated, with the source named.
- Push affiliate links over honest answers. Affiliate revenue is the byproduct, not the visible goal — see how we make money for the full disclosure.
- Pretend local AI is the answer for every workload. Cloud frontier models are still better at some things; the site says so directly when relevant.
Our test hardware
The active first-party benchmark fleet is named on the benchmark protocol page and on each benchmark row. Published owner-measured rows currently include the RTX 5080 desktop capture set where the public evidence package meets the repeat-run gate. Additional rigs are used for spot-checks and community-benchmark reproduction only when the row names the exact hardware. The discipline rule is clear: a verdict only counts as “measured” on the rig where it was run, and that rig is named on the benchmark row.
What this means honestly: the site has a growing first-party corpus, not blanket first-party coverage across every hardware page. Verdicts on hardware without owner-run evidence (RTX 4090, M3 Ultra, RX 7900 XTX, etc.) are labeled exactly as what they are: derived from vendor spec sheets, community benchmark submissions (editorially reviewed before publication), reproduced public sources, or computed fit math. Each row carries a provenance badge - Measured here, Source-backed, Community, Extrapolated, Estimated - and no row gets to claim a higher tier than its source supports.
Cross-vendor coverage (NVIDIA CUDA + AMD ROCm + Apple MLX/Metal) on editorial verdicts reflects published spec sheets and aggregated community benchmark runs, not direct first-party measurement on each platform. We’re honest about that gap; the long-term plan is to widen the first-party rig fleet as the editorial budget allows. The methodology behind every measurement is documented at /methodology.
How content is created
The bylined operator on every page is fred-oline. AI assistance is used for drafting, structuring, and maintenance scans — never for replacing operator judgment on buyer-decision content. The operator reviews and approves every piece of editorial output before it ships. See the editorial policy for the full process and the editorial philosophy for the trust-first principles that govern how the site weighs commercial pressure against reader interest.
How the site makes money
Affiliate links (Amazon Associates tag fredoline-20, plus first-class US retailers like Newegg and B&H for hardware) and display advertising. The site discloses these relationships on every page that carries affiliate-bearing links — see How we make money for the full disclosure including how editorial decisions are insulated from advertiser pressure.
Contact + corrections
For corrections, factual disagreements, hardware tip-offs, or partnership inquiries: contact page. The site logs every correction it ships in the changelog — operator transparency is the durability mechanism.