LifeArchitect Models Table: 1,000 Models With Their Training Receipts

⬅️ Back to Tools

LifeArchitect Models Table

What it isfilterable database of 1,000+ AI models with training stats and benchmark scores
Platformweb
Pricefree to browse; exports need a subscription
Linklifearchitect.ai/models-table

Leaderboards tell you which model wins. This table tells you what each model is made of. Dr Alan Thompson’s Models Table lists over a thousand models with parameter counts, training tokens, token-to-param ratios, architecture, and four benchmark scores side by side, updated hourly from public data. When I want to know whether a new release is actually big or just marketed big, this is where I look.

  1. The columns nobody else keeps are the reason to use it. Parameters and training tokens per model, the ratio between them color-coded against Chinchilla’s 20-to-1 rule, plus ALScore, Thompson’s quick power rating from the square root of params times tokens over 300. Arena scores tell you about vibes. These columns tell you about budgets.

  2. Benchmarks come with their ceilings printed on them, which is rare honesty. GPQA tops out around 90% and Humanity’s Last Exam around 25.6% on genuinely hard questions, with paper links beside both claims. A leaderboard that tells you where its own metrics stop meaning anything is one you can actually reason from.

  3. The filters answer real questions fast. Show only sub-30B models when you need something that runs locally. Split Chinese versus US labs, OpenAI versus Anthropic, diffusion models, or SOTA-at-launch only. Each filter is a different argument about the state of the field, one click away.

  4. The bubble visualizations are worth the visit alone. Planet-sized bubbles per model going back to April 2022, downloadable as PDFs with permission to reuse with citation. Watching five years of bubbles inflate is the fastest possible education in the scaling era, and the China-only and data-ratio charts sit beside them.

  5. The limitations: closed-lab numbers are estimates, since labs that do not disclose get inferred entries, and the whole edifice is one expert’s curation rather than an institution’s. CSV, JSON, and XLSX exports sit behind the Pro subscription, so programmatic users pay while browsers read free. And benchmark columns saturate exactly the way the page warns they do, which bounds what the table can tell you about the frontier.

Install & first run 

Open https://lifearchitect.ai/models-table/ in any browser.
Pick a filter: ≤30B for local runners, SOTA at launch for the frontier.

No account, no install. You should see the rankings view with benchmark leaders and the ceiling notes up top. For history, scroll to the bubbles visualizations and open any year’s PDF.

Need the data itself: exports and full Pro access require a subscription.

That is the whole paywall, stated on the page. Browsing, filtering, and the visualization PDFs stay free.

Worth your time if: you choose models by something other than vibes and want training stats beside benchmark scores. Skip it if you only need today’s arena winner, which simpler leaderboards answer faster.

Related TMFNK Content

Crepi il lupo! 🐺