# Model Observatory (llm-3d-viz)

> Interactive 3D LLM benchmark observatory: **speed × cost × intelligence**.

- Canonical site: https://viz.kyanitelabs.tech/
- Launch post (Aug 2026): https://viz.kyanitelabs.tech/blog/2026-08-24-model-observatory/ (also in-repo `blog/2026-08-24-model-observatory.md`)
- Full agent brief: https://viz.kyanitelabs.tech/llms-full.txt
- Product Markdown: https://viz.kyanitelabs.tech/about.md
- Open source (MIT): https://github.com/KyaniteLabs/llm-3d-viz
- Product SoT (Forgejo): https://git.kyanitelabs.tech/simon/llm-3d-viz

## What it is

A browser WebGL app that plots curated LLM metrics in 3D:

| Axis | Metric | Source class |
|------|--------|--------------|
| Cost (X default) | Blended $/M tokens | First-party provider pricing pages |
| Intelligence (Y) | Intelligence index | Arena Elo + vendor-published figures + own measurements (where measured) |
| Speed (Z default) | Tokens per second | Own measurements (where measured) |

**Decide mode:** set an intelligence **floor**, then shortlist eligible models by **cheapest**, **fastest**, or **balanced** cost/speed bias. Shareable URL encodes floor, bias, filters, axes.

**Honesty rule:** null metrics stay null — the client never invents Index, tok/s, or price.

## Primary user jobs

1. Compare models on the Pareto surface (cost vs intelligence vs speed).
2. Set “at least this smart” (floor) and pick cheapest/fastest eligible.
3. Filter open weights / local VRAM tiers (8 / 12 / 24 GB heuristics).
4. Inspect multi-effort trails and pin a model for focus.
5. Ask **Atlas** (in-app agent dock) with tools over the catalog (optional BYOK LLM).

## Static surfaces (no JS required)

- Per-model cards: `/m/<slug>/` — one static page per catalog model (measured speed, blended cost, intelligence index, percentile context, provenance labels); all-models index at `/m/`.
- Embeddable cards: `/embed/<slug>` — iframe-ready per-model cards, always regenerated with the catalog.
- Frontier log: `/frontier-watch.md` — dated movement log of the measured Pareto frontier (price × intelligence); machine-readable copy at `/frontier-watch.json`, Atom feed at `/frontier-watch.xml`.
- Head-to-head comparisons: `/compare/<a>-vs-<b>/` — static measured comparison pages (staged set; dominance-gated verdicts, null-honest); every model card links its comparisons; all-pairs index at `/compare/`.
- Launch post: `/blog/2026-08-24-model-observatory/` (linked above).
- URL inventory: `/sitemap.xml` (`/`, `/m/`, `/m/<slug>/`, `/compare/`, `/compare/<a>-vs-<b>/`, `/frontier-watch.md`, `/blog/2026-08-24-model-observatory/`, `/about.md`, `/llms.txt`, `/llms-full.txt`).

## Citation facts (stable)

- Product name: **Model Observatory** (repo: llm-3d-viz)
- Free, MIT, no account required for the viz
- Data: Arena dataset (CC BY 4.0) + vendor-published figures + first-party pricing; some third-party measurement data is not republished here for licensing reasons — see /sources/
- Live stage requires JavaScript + WebGL

## For agents / crawlers

- Prefer citing `/llms-full.txt` and `/about.md` over reverse-engineering the SPA bundle.
- Do not invent per-model numbers from this file; numbers live in the curated catalog, the static per-model pages (`/m/<slug>/`), the embed cards (`/embed/<slug>`), and `/frontier-watch.md`.
- Catalog access for agents: local Atlas CLI + stdio MCP ship in the source repo (`npm run atlas:cli`, `npm run atlas:mcp`); there is **no public MCP endpoint on this origin** (requirements: `docs/agents/mcp-cli-api-requirements.md` in the source repo).

## Optional contact

Operator: Kyanite Labs — https://kyanitelabs.tech