← Back to the work

MODEL SIGNAL / MODEL COMPARISON

Model Signal

Choose a task. Compare the models. See where the numbers came from.

Open Model Signal ↗
Model Signal showing task rankings, open-weight and proprietary filters, and published score ranges.
The current comparison interface. Open full size ↗

I wanted to compare models for the task I actually had.

I wanted a clearer way to compare models for the work I was doing. A writing score does not answer a coding question, and a ranking is hard to judge without its source. Model Signal puts the task first and keeps the evidence beside the result.

You can compare three models and check where their scores came from.

Choose a task, filter by model or provider and shortlist up to three models. The comparison keeps published values, score ranges and source links together. A separate library shows public-weight information and benchmark archives.

I update the data once a day and keep the original source dates.

I collect the comparison data once a day using n8n, which runs the scheduled updates. The site reads a saved copy of that data. That means visitors can browse and reload the comparisons without starting a new collection or paying for a model call.

Input validation protects the last good publication when an update is malformed. The interface retains each source’s dates and units, so an older benchmark does not look like a newly run test.

IMPLEMENTATION

How I built the comparison and its daily updates.

The comparison

Astro serves the page; React and TypeScript handle filters, shortlists and comparison states. Published results come from Arena, HELM and EvalPlus, with model metadata from Hugging Face.

  • Astro
  • React
  • TypeScript
  • n8n
  • Cloudflare KV
  • Zod
n8n
Scheduled collection and publication.
Zod
Validation of incoming datasets.
Cloudflare KV
Saved publications and a bundled fallback for browsing.
Vitest / Playwright
Data checks, interaction tests and responsive layout checks.