Zuvra

a quiet, private AI companion

Free AI chat that runs offline on your computer — fast on low-end PCs with no graphics card. No accounts. No cloud.

Get the engine

1.08B on-device model 131,072 context tokens 100% local & private Apache-2.0 model Free to use

First sight

Your first screen. Calm, empty, waiting.

The Zuvra screen

Why Zuvra

Six quiet luxuries — crafted, private, and already yours.

◈Privacy, built in

Chats live only on this device. One tap deletes — wiped from memory, screen, and disk. Truly gone.

★Watch it think

Reasoning unfolds live in its own panel. Follow it, read the answer, or stop it mid-thought.

↯Answers, as they happen

Words stream live; your tab announces each reply.

◔Every token, accounted for

Each reply shows its token cost. Hover for the breakdown.

☰Your past, at hand

Search, rename, organize — in a drawer that never interrupts.

▣Code highlighting

Code highlighting for 23 languages and formats. Python · JavaScript · TypeScript · Java · C · C++ · C# · Go · Rust · PHP · Ruby · Swift · Kotlin · SQL · HTML · CSS · Shell · JSON · YAML · TOML · INI · Dockerfile · Diff. Copy any block with one tap — HTML and JavaScript open as runnable previews.

Run it on your machine in 3 steps

The chat talks only to your computer. Its engine is the free MiniCPM Desk Pet — install it, keep it running, done.

1Get the engine

Download the official desktop app (v0.10.0). One click starts the download from GitHub Releases:

Download the engine

Detecting your device…

All versions: official releases page.

2Install & run it

Install the program, then open it. Keep it open while you use Zuvra.

3Open your chat

Download Zuvra above, open Zuvra.html — no internet needed. Nothing is ever uploaded.

Requirement: the chat file needs MiniCPM Desk Pet running (the desktop app and local engine). Without it you will see “Offline — open MiniCPM Desk Pet and retry.” The chat itself never contacts any external server.

Take everything with you: one-click copy everywhere, runnable previews, searchable history. On your device, free.

Answers begin at onceThinking first, live — never a blank screen.

Light on your device1.08B Q8 — sips CPU and RAM on ordinary computers.

Zero network waitingOn-device: no queues, no accounts. Speed depends on your hardware.

Small but serious1B-class SOTA (per OpenBMB), hybrid reasoning — more with less.

Questions, answered

What you should know

Which computers can run Zuvra?

Any ordinary Windows or macOS computer. Zuvra runs on MiniCPM5-1B, a 1.08-billion-parameter model small enough to stay fast on low-end PCs with no graphics card, using little CPU and RAM.

Does Zuvra work without internet?

Yes. After a one-time install of the free MiniCPM Desk Pet engine, everything runs on your computer. No connection needed and nothing is uploaded.

Is my data private?

Chats stay on your device. One tap deletes a conversation from memory, screen, and disk. There are no accounts and no cloud servers holding your data.

How much does Zuvra cost?

Free. The Zuvra interface is free, and the MiniCPM engine is free and open-source under the Apache-2.0 license.

What can't Zuvra do?

The model is text-only: no images, audio, or video. Zuvra also needs the MiniCPM Desk Pet app running — without it, the chat shows an offline notice.

Built with MiniCPM

Powered inside by MiniCPM 5

Zuvra's inner engine is MiniCPM5-1B — a 1.08B on-device model by OpenBMB, running locally as a GGUF quant. 131,072 tokens of memory in a 1B body. Text only — it understands written words, not images, audio, or video.

1.08Bparameters 131Kcontext length Q8_0local GGUF quant Thinkhybrid reasoning Textonly · no images

License & attribution. MiniCPM5-1B © OpenBMB — this repository and the MiniCPM model weights are released under the Apache License, Version 2.0. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0. Model: openbmb/MiniCPM5-1B · Source: OpenBMB/MiniCPM · Paper: MiniCPM Tech Report · Desk Pet app: MiniCPM-Desk-Pet (AGPL-3.0). Built by ModelBest (Beijing) with Tsinghua NLP and the OpenBMB open-source community. The Zuvra interface is a separate original work; the model weights run locally on-device and this page distributes no model files.

Model card Apache-2.0 text