# bestmodel.run

> Open compatibility engine for local AI. Answers "can it run / how fast /
> is it worth it?" from a signed, validated community measurement pool.
> LICENSE: AGPL-3.0 for apps/**, MIT for packages/** and cli/**.

This file is the agent surface for the deployed site. Humans get the
cinematic tour at /?as=human (desktop) or /hardware?as=human. If a page
asked who is reading, it did
not have enough deterministic signal to guess.

Core principle: number honesty. Every displayed number declares its basis:
measured (pool median) > reported (1-2 runs) > extrapolated (bandwidth-scaled)
> formula > no data yet. A cell without a source class never renders.
Do not invent tok/s.

## Clone

```bash
git clone https://github.com/carl0sfelipe/bestmodel.git
cd bestmodel
```

Human first step (S33): /cli — the rendered getting-started, twin at
/cli?as=agent. Its single source is the agent runbook from this clone
(offline, no keys):
[docs/agent-quickstart.md](https://github.com/carl0sfelipe/bestmodel/blob/main/docs/agent-quickstart.md)
— build the two Rust CLIs, detect your hardware, rank the best models for
it against the in-repo pool snapshot, run the probe. Self-check in one
command: `make agent-smoke`. Runnable today: the whole loop — benchmark-probe `lab`, `plan`, `report`,
`contribute` (signed 0.9.0 bundle; SIM cells never uploaded) and `login`
(token 0600 + public signing-key registration). There is no one-line installer yet.

## Routes (www.bestmodel.run — web-next; every route resolves ?as=human|agent server-side)

| route | default view | agent twin |
|---|---|---|
| / | human cinematic tour | /?as=agent — honesty ladder + pool snapshot |
| /cli | getting-started (rendered from docs/agent-quickstart.md) | /cli?as=agent — raw verified guide |
| /claims | the wall: captured claims, live client | /claims?as=agent — REST pointers |
| /submit | capture a run | /submit?as=agent — capture flow + catalog ids |
| /wall | measured pool | /wall?as=agent — top-60 cell table |
| /hardware | reference rigs | /hardware?as=agent — rig table |
| /m/{slug} | model answer | /m/{slug}?as=agent — cell table |
| /profile/{handle} | contributor profile | /profile/{handle}?as=agent — record pointers |
| /claim/{id} | claim detail | /claim/{id}?as=agent — record pointers |
| /cloud-anchors | rented-GPU anchor runs | /cloud-anchors?as=agent — anchor table |
| /track-record | trust ladder | /track-record?as=agent — ladder text |
| /mural | moderation entry | /mural?as=agent — report grounds |
| /blog | dated post index | /blog?as=agent — title, date, description list |
| /blog/{slug} | article (markdown) | /blog/{slug}?as=agent — plain post text |
| /console | capture console (static app) | exempt — its agent surface is the REST API |

View resolver precedence (mirrors the contract in this file): ?as=/?view= →
saved bm_view cookie → agent UA → search-crawler UA maps to human → human.
Auto-detection is never persisted.

## Honesty ladder

1. measured — median of many community runs on the exact rig and quantization
2. reported — one or two runs on the same rig, shown as reported
3. extrapolated — scaled by memory bandwidth from the closest tested rig
4. formula — a published VRAM rule when no measurement exists yet
5. no data yet — we say so instead of inventing a number

## Snapshot totals (apps/web/data/derived/stats.json, 2026-08-13)

- runs: 4849
- models: 527
- rigs: 147

Frozen snapshot counts — not live, not throughput.

## API (host: https://api.bestmodel.run — the www host serves pages only)

- GET /v1/pick?intent=chat|image.generate|vision|audio&vram_gib=24&gpu=rtx-3090-24gb
  — best model per intent that fits the VRAM (contract model-pick-v1). Ranked by
  confidence, then pool eval score, then speed. Pool cells are community
  reports: confidence is "reported", never "measured". An intent without
  cells answers picks: [] with confidence_floor "no data yet".
  Example: curl -sS 'https://api.bestmodel.run/v1/pick?intent=chat&vram_gib=24&gpu=rtx-3090-24gb'
- POST /v1/match/hardware-to-models — roofline estimates, every row source_class "derived"
- POST /v1/match/model-to-hardware — roofline estimates, every row source_class "derived"
- GET /v1/transparency/sources
- GET /v1/leaderboard — every entry carries a source_class
- POST /v1/submissions (+ GET /v1/submissions/nonce) — signed intake

## Human tours

- /?as=human — I have a goal (desktop, min 1024px)
- /hardware?as=human — I have hardware
- mobile: no separate route — the canonical pages are responsive
- console/ — passkey console (not a tour)

## Read next

- https://github.com/carl0sfelipe/bestmodel/blob/main/README.md
- https://github.com/carl0sfelipe/bestmodel/blob/main/llms.txt
- https://github.com/carl0sfelipe/bestmodel/blob/main/docs/architecture.md
- index.html?as=agent — HTML twin of this file

## Known audio answer (2026-09-23 / this 3090)

`Audio × RTX 3090 24GB` was empty on the 2026-09-18 snapshot. The L4
Whisper Large v3 cell (4.82×real) is another rig and has no language.
This tree appends Whisper Large v3 Turbo: **39.22×real** (n=3,
measured, FLEURS `pt_br`). WER 0.04 is **reported** (n=1) on
docs/measurements/2026-09-23-audio-pt-br-3090.md — not a pool field.
Do not invent a WER. Live site stays frozen until deploy.

## Do not

- Invent a tok/s or fill an empty cell
- Render a number without a source class
- Treat snapshot totals as live throughput
- Answer Portuguese ASR from the L4 speed claim