You ask.
In plain words. No prompt tricks, no settings.
Write a press release about Muxxy.
IT WROTEMuxxy is a video infrastructure platform that enables creators and companies to build, deliver, and monetize video at scale. Muxxy's platform is used by thousands of companies and creators worldwide.

Meet Nado In private pilot with design partners
A private AI that runs on your own computer, with plug-in knowledge. It shows its sources or says it does not know.
See it answer How it measures upIn plain words. No prompt tricks, no settings.
Nado keeps cartridges: knowledge and saved procedures, each with a contract. It picks the one that matches your request, or none at all.
Each number in the answer has to appear in the source it cites. If one does not, the answer never reaches you.
Citrusdal is a town of 7,177 people as of 2011, located in the Olifants River Valley in the Western Cape province of South Africa [1]
Where it looked, what it used, and the check it passed. Every time.
REQUESTShort cited article on Citrusdal's census population.
ANSWERCitrusdal is a town of 7,177 people as of 2011, located in the Olifants River Valley in the Western Cape province of South Africa [1]
Real output from our earlier 27 Sep 2026 test of procedure cartridges.

Press release: Muxxy names Thabo Nkosi Head of Product, effective 1 November, dateline Cape Town.
WHYIts saved press-release procedure announces company milestones, not new hires. So it said so, instead of guessing.
Hot-swappable Knowledge cartridges
A cartridge is a folder of your own documents, turned into a searchable, cited index with one command. Plug it in, and Nado answers from it, citing the file and the place in it.
One command turns a folder of your documents into a searchable, cited index.
Add, update or remove a cartridge and Nado uses the change straight away. No restart, no retraining. The model itself never changes.
Nado picks the cartridge whose name and description fit the question, and cites the file and the location.
Removing a cartridge deletes only Nado's index, never your documents.
Nado help comes built in. Nado ships with its own documentation as a cartridge, so you can ask Nado how to use Nado.
New in the kit Three things Nado does now
Same rule as everything else: you can see where each answer came from, and Nado says so when it can't back one up.
Start with “Remember:”, “remember that…” or “keep in mind…” and Nado saves the fact as a note on your computer. Questions that touch it are answered from the note, and say so. “Forget…” removes one; “show my notes” lists them all.
The notes are a plain file you can open and edit, and a cited cartridge (my-notes). Nothing is sent anywhere.
Nado's working memory is about 8,000 tokens per conversation. We won't pretend otherwise. So instead of keeping only the end of a long attachment, Nado splits it into located parts (page numbers for PDFs) and brings in the parts that fit each question, plus the opening. Answers can cite the page.
Ask about something current and Nado writes a few searches, reads the best pages and answers only from them: a numbered citation after every claim, links listed. If the pages don't cover it, it says so. Ask for research and it runs two rounds (search, read, find what's missing, search again) and writes a structured, cited report.
Live timings on the laptop. Only a few short search queries and the page requests leave your computer; your files, cartridges and the conversation never do. Offline mode turns all web access off.
Your knowledge doesn't need a giant context window. It lives in cartridges, and long documents are searched, not truncated.
Five reasons, none of which need you to take our word for it.
Your files and documents stay on your computer. What does leave is small: the licence check, and short search queries and page requests when Nado looks something up on the web. Offline mode switches all web access off.
A Mac with Apple silicon today; Windows 10 and 11 (x64) is new and in testing. It listens only on 127.0.0.1.
Checked answers come with where they came from. When Nado cannot back an answer, it declines instead of guessing.
Hot-swappable cartridges from your own folders, cited on every answer and live from the next question (see above). Bring your Claude MCP servers, skills and instructions across with nado-import-claude.
Nado runs on hardware you already own, so there is no bill per token. Ask ten questions or ten thousand.
Tap a request. Nado replays exactly what it did with it in our earlier 27 Sep test, receipt and all.

Replays of recorded runs from our earlier 27 Sep 2026 test of procedure cartridges, shown as they happened. Not a live model: Nado is in private pilot, and there is no public chat.
GEN14 Release gate · 12 Oct 2026
Within 10 points of the frontier. On your own laptop.
Sealed test, 1,240 unseen tasks, all 12 areas. Your files never leave your computer.
areas within 10 points of the frontier, on a sealed, pre-registered test of 1,240 unseen items
| Area | Nado | Frontier | Gap |
|---|---|---|---|
| Abstain | 95 | 100 | −5 |
| Code | 97 | 100 | −3 |
| Data | 100 | 100 | 0 |
| Instruction | 88 | 91 | −3 |
| Knowledge | 95 | 99 | −4 |
| Logic | 100 | 100 | 0 |
| Math | 99 | 100 | −1 |
| Reading | 84 | 73 | +11 |
| Tools | 94 | 96 | −2 |
| Writing | 88 | 72 | +16 |
| Audio* | 95 | 97 | −2 |
| Image | 99 | 99 | 0 |
Within 10 points, not equal: Nado trails the frontier by 1–5 points in 7 of 12 areas, ties in 3 and leads in reading and writing. Items were written by DeepSeek and are not public benchmarks; public-benchmark results are not yet published.
GEN14 also ran Gemma 4 E4B alone, as a reference. Alone, it already scores high on knowledge (98), code (98), math (100) and reading (86). Nado's system adds most on instruction following and business writing, plus data (95 → 100) and abstaining (90 → 95). It does not turn a small model into a frontier model. It makes a strong open model reliable, private and useful on your own computer.
| Area | Gemma alone | Nado | Frontier |
|---|---|---|---|
| Instruction | 71 | 88 | 91 |
| Writing | 50 | 88 | 72 |
How we tested. GEN14 is a sealed set of 1,240 items neither contestant had seen: 1,000 text items across 10 areas, written by DeepSeek V4 Pro (neither contestant's model family), and 240 image and audio items. Each item was graded mechanically or by a fixed judge. The rule was written down first: non-inferiority within 10 percentage points in every one of the 12 areas, on paired 95% bounds. The frontier was Claude Opus 5.5 for text and images, run as Claude Code subagents with no tools, and Gemini 3.1 Pro for audio. Gemma 4 E4B alone ran as a reference. Nado, Gemma 4 E4B inside the Nado system, passed in 12 of 12 areas. This was its release gate, passed on 12 Oct 2026.
Earlier test 27 Sep 2026 · procedure cartridges v5

In an earlier test of procedure cartridges, when Nado answered it was right 48 times out of 50. We wrote the requests and froze them before the code ran, and the model behind every answer ran on one laptop.
Every Saturday at eight-thirty in the morning, generate the weekly usage report.
WHAT HAPPENEDIt booked the job for 8:00. Its time reader does not yet understand minutes written as words. That is the next fix.
Procedure cartridges, v5, run 27 Sep 2026 on two wordings of each request, on a different model (K2). Prediction and failure condition written before the run. A separate, earlier experiment: a small sample of business tasks, not a comparison with frontier models, and not part of GEN14.
Want to test it? Request a pilot place. Already a design partner? Sign in at muxxy.ai/partner with your email and invite code for downloads and docs.
Muxxy is an AI lab. We are not Mux, the video company, and not connected to it. Now you know why we say so.