RunLocal

The independent guide to running AI locally.

Detect your GPU in the browser → see which open-weights models actually fit your VRAM → calculate what self-hosting saves versus commercial APIs. Built as a content-first directory site, not just a benchmark table.

Status: early prototype. Benchmark scores are a snapshot from local.ai (Exo Labs, 2026-08). Quant sizes and savings are estimates. Pricing values are placeholders until a live pricing pipeline is added.

Features

• Browser GPU detection — WebGPU GPUAdapterInfo + WebGL WEBGL_debug_renderer_info, with manual hardware selection as a fallback for Safari / old cards / remote desktops.
• VRAM-aware recommendations — maps your chip to a hardware database and recommends the best quantization that fits your memory budget.
• Self-hosting savings calculator — uses an open model's own API price when it exists, otherwise its closest commercial equivalent; compares API spend against electricity + hardware amortization.
• Model directory + detail pages — 44 seeded models, statically generated with JSON-LD, sitemap, internal links and per-page VRAM/quant/cost tools.
• Zero tracking, zero accounts — detection happens entirely in the visitor's browser.

Quick start

npm install
npm run dev      # http://localhost:3000
npm run build    # static export to out/ (Cloudflare Pages ready)

Deploy to Cloudflare Pages

Option A — Git integration (recommended): in the Cloudflare dashboard, create a Pages project connected to this repo with:

• Build command: npm run build
• Output directory: out
• Environment variable: NEXT_PUBLIC_SITE_URL = your production URL

Option B — Wrangler CLI:

npx wrangler login
npx wrangler pages deploy out --project-name runlocal

Option C — GitHub Actions: the included workflow (.github/workflows/deploy.yml) rebuilds and deploys on every push to main and daily at 04:23 UTC. Add these secrets/vars to the repo settings:

• CLOUDFLARE_API_TOKEN (Pages edit permission)
• CLOUDFLARE_ACCOUNT_ID
• Repository variable NEXT_PUBLIC_SITE_URL

Data pipeline

The directory seed lives in lib/data.ts; live data lives in data/generated/.

npm run ingest            # HuggingFace metadata + OpenRouter pricing snapshots
npm run ingest:models     # models only (license, downloads, GGUF file sizes)
npm run ingest:pricing    # pricing only

The scripts are polite to rate limits (retry with backoff, keep previous data on partial failure) and use HTTPS_PROXY automatically when it is set. Commit data/generated/ so a fresh clone builds without network access.

Repo layout

app/                    Next.js App Router (home, /models, /models/[slug])
components/             GPU detect, recommendations, savings calculator, quant pills
lib/                    data seed, detection, ROI math, pricing, quants, SEO helpers
prototype/              original single-file vanilla-JS prototype
content/SEO_CONTENT_PLAN.md   50-page content roadmap with keyword matrix
docs/PLAN.md           product & monetization strategy
docs/RESEARCH-local-ai.md     notes from reverse-engineering local.ai
docs/PLAN.zh-CN.md     中文版产品方案

Product strategy (TL;DR)

docs/PLAN.md has the full version:

1. The detection tool is the hook (shareable, retentive), not the traffic engine.
2. Model pages, VRAM guides, comparisons and news are the SEO engine.
3. Every piece of content ends at a decision page (what to run / what to buy).
4. Monetization order: hardware affiliate → cloud-GPU affiliate → sponsored newsletter → ads at scale → data/API licensing later.

Roadmap

• Pricing pipeline: daily snapshots from provider pages / OpenRouter
• Model metadata pipeline: HuggingFace API + Ollama library (license, context, files, sizes)
• First 20 model detail pages rewritten with real metadata
• 10 guides + 10 comparisons (see content plan)
• Search/filter on the directory
• Hardware detail pages and comparison pages (affiliate-ready)
• Community benchmark submissions

License

MIT