Model Beats

The definitive hub for LLM rankings, benchmarks, evaluations, model comparisons — and the only one that also answers “can my hardware run it, at which quant, how fast?”

Model Beats collapses the four-tab workflow — rank it, verify it, size it, run it — into one provenance-honest, deep-linkable tool: every major LLM release from GPT-3 (2020) to today — 597 models across 86 orgs, 171 benchmarks with published normalization bounds, version lineage, API pricing, quantizations, and a graded hardware-fit engine. System design in ARCHITECTURE.md (§13 reconciles it with the committed design handoff); every locked decision and contract in docs/DECISIONS.md.

Screens

/ dashboard (overview · releases · benchmarks tabs) — /rankings (+/{category}) dense sortable table — /models faceted explorer with runs-on-my-GPU — /models/{slug} detail with benchmark provenance, family lineage and the Run-it-locally card — /compare?m=a,b,c specs/benchmarks/radar with saved comparisons — /hardware graded fit verdicts by profile or by model — /benchmarks/{slug} leaderboards with provenance badges and distributions — org/family hubs — /methodology — full SEO surface (sitemap, JSON-LD, canonical).

URL is the state everywhere — every filter, sort, tab and comparison is a shareable link. Dark (default) + light themes from the design-token palette.

Architecture in one paragraph

One Cloudflare Worker serves static assets + TanStack Start SSR + two typed server functions. There is no database: bun run build-catalog computes everything from the curated data/** files into two immutable, content-versioned artifacts — a headline catalog snapshot and a per-model detail map — which are bundled into the Worker at build time and served from the edge cache. The catalog ships to the client once and powers all list/filter/rank interactions through pure shared selectors. Scores are computed at build time (validate → derive → build catalog); a data change is a new build, hence a new content version, hence fresh caches — nothing to purge. The rating formula (Frontier Elo — Bradley-Terry over pairwise benchmark battles, D21), hardware-fit thresholds and snapshot schema are golden-tested contracts (packages/shared).

Development

Prereqs: Bun ≥ 1.3 and Node ≥ 24 (wrangler/vite/vitest run under Node — never under the Bun runtime). No Cloudflare account needed to develop.

bun install
bun run build-catalog   # validate → derive → build the catalog + model-detail artifacts
bun run dev             # http://localhost:3000

Command	What it does
bun run build-catalog	validate → derive → build the bundled serving artifacts (run before typecheck/build/dev)
bun run ci	build-catalog → typecheck → lint → unit tests → build → perf budgets (exactly what CI runs)
bun run e2e	build the artifacts, then Playwright against the built preview in workerd
bun run validate-data	Zod + cross-file integrity gates over data/
bun run derive	recompute data/derived/scores.json (committed, reviewable)
bun run deploy:production	credential-gated real deploy (also what CD runs on merge) — see docs/DEPLOY.md

Repository layout

apps/web/          TanStack Start app (routes, screens, server fns, e2e specs)
packages/shared/   Zod schemas · scoring engine · hardware-fit engine · selectors · formatters · contracts
data/              curated dataset — the repo is the CMS (see CONTRIBUTING.md)
scripts/           validate / derive / build-catalog / budgets / deploy
docs/              design handoff · DECISIONS.md · DEPLOY.md

Data & contributions

The dataset is curated, versioned and validated in git — numbers are point-in-time approximations with per-result provenance, disclosed on /methodology. To add a model or benchmark (or fix a score), see CONTRIBUTING.md: CI enforces the gates, and the derived-scores diff in your PR shows exactly how rankings move.