Your personal news anchor.
On your machine.
Tell Pepper what you care about, in plain words. She researches it around the clock, writes the bulletin with an on-device model, and presents it from her own studio — with her own voice. No cloud, no accounts, no telemetry.
Open source, AGPL-3.0 · Benchmarked, not vibes · Full on-device brain on Apple Silicon
Install — thirty seconds to air
curl -fsSL pepper.software | shpepper startPrefer npm? npm i -g @pepperchan/pepper · Node 18.17+ · everything lives in ~/.pepper · her voice tier: pepper voice install
Tell her your beats
"Watch open-source robotics and GPU prices for me." Plain words in her studio — she works out the rest and starts sweeping the wire every fifteen minutes.
She writes the bulletin
An on-device model turns the wire into a sourced broadcast — vendor claims labeled, rumors kept off the air, nothing invented. Your reading interests never leave the machine.
She goes on air
A 3D newsroom in your browser: chyron, ticker, her designed voice, deep-dive reports on demand — and a one-command export that puts your station on the public web.
Benchmarked, not vibes
Best in her weight class — and honest about the ceiling above it.
MoltBench scores wire-to-report work: grounding, attribution, and adjudication of deliberately contaminated bundles, under blind judges. Pepper's first model failed it — that failure became the release gate. Her current desk model cleared it: +23 grounding points over its own untuned base, 10/12 blind wins against Qwen3-4B-2507, 11/12 against Qwen2.5-7B. Models 6× and 13× her size still out-broadcast her — 10/12 and 11/12 — and those rows are published too, because a leaderboard that hides its ceiling isn't one. The judge column is the claim; regex grounding is only the screening metric (it rewards terseness — a lesson this benchmark learned in public). Protocol and bundles →
| Model | Blind judges vs pepper-desk | Regex grounding (screening) |
|---|---|---|
| pepper-desk-e2b (2B active) | — | 88.6% |
| Qwen3-4B-Instruct-2507 (same class) | she wins 10 / 12 | 88.3% |
| Qwen2.5-7B-Instruct | she wins 11 / 12 | 77.5% |
| gemma-4-12b (her family, 6× the size) | she loses 2 / 12 | 74.3% |
| Qwen3.8-27B (13× the size) | she loses 1 / 12 | 80.9% |
| gemma-4-e2b (her untuned base) | — | 65.2% |
Run a station of your own
Three ways to put a Pepper on the web.
The reference station at pepper.watch — served static from Cloudflare Pages, her voice included.
Run Pepper locally, then pepper export → npx wrangler pages deploy pepper-site. Your beats, your bulletins, your voice files — a complete station from your own desk.
Deploy the station template straight to your Cloudflare account, then feed it from your desk whenever you like.