ModelFit

Independent model-to-hardware planning site. The launch UI is branded ModelFit through NEXT_PUBLIC_SITE_NAME; the final brand and domain are environment-driven.

The product direction is inspired by RunMyLLM’s model/hardware decision flow, with additional production-serving inputs, internal hardware profiles, and a selective compute handoff after the fit result.

Included

• Interactive weight, KV-cache, and runtime-memory estimator
• Quantization, context, concurrency, cache precision, and GPU-count controls
• GPU-fit recommendations with hourly Flatkey price-index data where covered and labeled reviewed references elsewhere
• Model directory plus 11 indexable model requirement pages
• GPU hardware directory plus 9 model-fit profile pages
• Six substantial VRAM, quantization, multi-GPU, model-selection, and hardware-comparison guides with FAQs
• Copyable vLLM command templates with explicit checkpoint and backend caveats
• Published formula and limitations page
• Per-page canonical metadata, structured data, sitemap, robots policy, standalone Next.js build, and Docker image
• One result-level Flatkey Compute CTA on eligible calculator, model-fit, and hardware-fit pages, with model and hardware UTM attribution
• Daily model-configuration refreshes from official Hugging Face repositories
• Consent-gated GA4 acquisition and calculator-funnel measurement

Local development

bun install
bun run dev -- --port 3001

Open http://localhost:3001.

Production build

bun run typecheck
bun run build
PORT=3001 node .next/standalone/server.js

Or build and run the included Dockerfile with -e PORT=3001 -p 3001:3001.

Refresh model data

bun run data:refresh
bun run test
bun run typecheck
bun run build

data/model-configs.json is a versioned, auditable snapshot. Public model configs update automatically each day; gated or temporarily unavailable configs retain their last-known-good values. Collector and validation code lives under scripts/.

See DATA.md for the data contract, formula, and fallback policy.

Configuration

Copy .env.example to .env.local and set:

• NEXT_PUBLIC_SITE_URL: canonical production origin
• NEXT_PUBLIC_SITE_NAME: final public brand
• NEXT_PUBLIC_FLATKEY_COMPUTE_URL: Flatkey Compute product entry (defaults to https://console.flatkey.ai/compute/market)
• NEXT_PUBLIC_GA_MEASUREMENT_ID: GA4 web stream measurement ID
• FLATKEY_PRICE_INDEX_URL: hourly GPU price-index API (defaults to Flatkey's public index)
• HUGGING_FACE_TOKEN: optional read token for gated official model configs

GA4 never initializes until flatkey_analytics_consent=granted is present in local storage. It measures acquisition, landing-page behavior, calculator progression, and the Flatkey Compute handoff for channel reporting.

Estimation scope

The current engine estimates inference memory. Results are labeled calculated estimates and are not benchmark guarantees. Measured runtime profiles should be stored separately with exact model revision, quantization, runtime version, GPU SKU/form factor, context, concurrency, and launch flags.

License

MIT