emu — AI For The Rest Of Us
emu is a local-first, multimodal AI desktop application that runs state-of-the-art AI models entirely on your own computer. Chat, dictate by voice, have replies read aloud, read documents and ask about images — all with a clean, minimal interface, and with no cloud account, no telemetry, and no network calls for inference.
A product of CloudKites AI Lab (CloudKites Pty Ltd), Sydney, Australia.
💚 Free to download, use, and share. emu is distributed free of charge. You may use it and redistribute the complete, unmodified official package for free (no selling, no paid bundling). See License below.
🎉 Latest release: emu 3.2.2 — now native on macOS, Windows and Linux. The previous edition, emu 2.0, is still available.
📘 New to emu? Start with the User Guide (PDF). It walks you through installing emu on Mac, Windows or Linux, choosing the right model for your computer, and using every feature step by step, with screenshots. No technical knowledge needed.
⬇️ Download emu 3.2 — pick your computer
The same app, with the same interface, on all three operating systems — native on each, with no runtime dependencies (nothing else to install).
| Your computer | Download | Install |
|---|---|---|
| 🍎 Mac — Apple silicon (M-series), macOS 13+ | emu-3.2.2-macos-arm64.dmg | Open it and drag emu to Applications |
| 🪟 Windows 10/11 (64-bit) | emu-3.2.2-windows-x64.zip | Unzip anywhere and double-click emu.exe — no installer |
| 🐧 Linux (x86-64) | emu-3.2.2-linux-x64.tar.gz | Extract, then ./install.sh (or run ./emu in place) |
Every file has a matching .sha256, and
SHA256SUMS
lists them all.
First launch.
- Mac: if macOS asks whether to open an app downloaded from the internet, choose Open (or right-click emu › Open).
- Windows:
emu.exeis code-signed by CLOUDKITES PTY LTD. SmartScreen may still warn about a brand-new release — click More info, check the publisher, then Run anyway. - Linux:
./install.shinstalls for you only (no sudo):~/.local/bin/emu, an applications-menu entry and an icon../install.sh --removeuninstalls. Install zenity or kdialog for Open/Save dialogs.
On first start emu shows Get the model and downloads and verifies (SHA-256) its
chat model, with resume if the download is interrupted — or Choose file… to
use a .nbq you already have. Voice, document and image models are downloaded the
first time you use each feature. After that, everything runs offline.
📘 How do I use it? Read the User Guide (PDF): step-by-step instructions with screenshots for every feature.
📄 Need help choosing or installing? See the Download & Installation Guide (PDF).
🎉 What's new in 3.2.2
- Every language draws — Chinese, Japanese, Korean, Arabic, Persian, Urdu, Hebrew, Hindi and the other Indic scripts, Thai, Khmer, Myanmar, Tibetan, Georgian, Armenian, Amharic, emoji and dozens more: no more empty boxes. Arabic and Indic letters join and combine correctly, and right-to-left text (Arabic, Hebrew, Persian, Urdu) reads right to left, with numbers and English words in the right order.
- A new, illustrated user guide (USER_GUIDE.pdf): installing, choosing a model for your computer, and every feature step by step, with screenshots.
- Fixes: long notifications wrap instead of running off the window; the delete-chat dialog wraps its text; the download progress bar no longer covers the line above it; a rare drawing bug after deleting a chat is fixed.
🎉 What's new in 3.2.1
- Maths, typeset — formulas in replies (fractions, roots, sums and integrals, matrices, Greek…) show as real maths instead of raw TeX, with every model.
- Tooltips on the buttons and clearer icons: a paperclip to attach, a framed T to read text from an image or PDF.
- Downloads check free disk space first, and say so plainly when the disk fills, then carry on where they stopped once there is room.
- Fixed: a speech or OCR pack in the models folder was mistaken for a chat model.
🎉 What's new in 3.2
- Windows and Linux editions of emu 3, next to the Mac edition — one app, one interface, no Qt, no Python, no installer needed on Windows.
- Chat on the processor on Windows and Linux: Qwen3.5 9B · 4-bit by default, or the lighter, faster Qwen3.5 4B · 4-bit ("Download faster…").
- Voice, documents and images on every OS — dictation (Qwen3-ASR, 30 languages), four speech engines with Vietnamese voices and voice cloning, every clip saved, PP-OCRv6 / Unlimited-OCR document reading, and image questions with Qwen3-VL.
- Native on each system — Ctrl shortcuts on Windows and Linux (⌘ on the Mac), input methods (Vietnamese Telex/VNI/Unikey, Chinese, Japanese…), high-DPI displays, dark/light theme following the system.
💬 The chat model on each computer
| Mac | Windows & Linux | |
|---|---|---|
| Model | Qwen3.8 27B · 4-bit | Qwen3.5 9B · 4-bit (default) or Qwen3.5 4B · 4-bit |
| Runs on | The Mac's GPU, with lossless speculative decoding (DFlash2 block drafting, or the model's own MTP head) — exactly the answers of plain decoding, only faster | The processor (CPU) |
| Speed | The fastest emu | Laptop (Intel Core i7-13800H, 32 GB): 9B ≈ 2.6 tokens/s, 4B ≈ 5.6 tokens/s · 32-core desktop: 9B ≈ 8 tokens/s |
On Windows and Linux, long prompts (a pasted or scanned document) take noticeably longer to start answering — tens of seconds for a page. On an older or low-memory PC, choose the 4B. Thinking (Off / Brief / Standard / Deep) is available on every edition.
📦 Model packs
Each model is one self-contained .nbq file (weights, tokenizer and helpers). The
app downloads the ones you need — you don't have to fetch them by hand.
Chat (emu 3.2)
| File | What | Size |
|---|---|---|
| qwen3.8-27b-q4km.nbq | Mac. Qwen3.8 27B (Unsloth UD-Q4_K_M, with its MTP head) + the DFlash2 draft model + the tokenizer, in one file | 17.6 GB |
| qwen3.5-9b-q4km.nbq | Windows & Linux, default. Qwen3.5 9B (Q4_K_M) + tokenizer | 5.7 GB |
| qwen3.5-4b-q4km.nbq | Windows & Linux, lighter and faster. Qwen3.5 4B (Q4_K_M) + tokenizer | 2.8 GB |
Voice, documents and images (emu 3.x, all platforms)
| File | What | Size |
|---|---|---|
qwen3-asr-0.6b.nbq |
Dictation — Qwen3-ASR 0.6B, 30 languages | 1.9 GB |
pocket-tts-en.nbq |
Speech — Pocket TTS, English, fastest | 0.26 GB |
qwen3-tts-0.6b.nbq |
Speech — Qwen3-TTS, multilingual, 9 voices | 2.5 GB |
gwen-tts-vn-0.6b.nbq |
Speech — gwen-tts, Vietnamese, 9 voices | 2.5 GB |
vieneu-tts.nbq |
Speech — VieNeu-TTS, Vietnamese + English, voice cloning | 0.4 GB |
ppocr-v6.nbq |
Documents — PP-OCRv6, fast; text and tables | 0.16 GB |
unlimited-ocr.nbq |
Documents — Unlimited-OCR, most accurate; layout, tables, maths | 6.7 GB |
qwen3vl-2b-cpu.nbq |
Images — Qwen3-VL 2B (Windows, Linux and Mac) | 4.8 GB |
qwen3vl-8b-cuda.nbq |
Images — Qwen3-VL 8B (Mac) | 6.7 GB |
SHA-256 of the chat packs (also in SHA256SUMS):
qwen3.8-27b-q4km.nbq—58e4798b003c49f7e7dd15df2d5ee33921bee6357137e46be3ceed9b08fa6646qwen3.5-9b-q4km.nbq—7659c84c636c1dc2e5a8d3b8443c64c6c7f29281fea8e3281e96e8c74ef67764qwen3.5-4b-q4km.nbq—d5c3411cd851c6adcbd670ee5c0f9cbe37e4cf7b40deedb549259615e7834fb1
The emu 2.0 packs
qwen3.5-9b-q4_k_m.nbqandqwen3.5-4b-q4_k_m.nbq(note the underscores) are a different format and do not load in emu 3 — use the…-q4km.nbqpacks above.
🖥️ Requirements (emu 3.2)
| Mac | Windows | Linux | |
|---|---|---|---|
| System | Apple silicon (M-series), macOS 13 or later | Windows 10 or 11, 64-bit | x86-64, glibc 2.31+ (Ubuntu 20.04, Debian 11, Fedora 32 or newer); X11, or Wayland through XWayland |
| Processor | — | AVX2 — Intel 4th-gen Core (Haswell) or newer, AMD Ryzen or newer | AVX2 |
| Memory | 32 GB recommended, 24 GB minimum | 16 GB for the 9B (8 GB with the 4B) | 16 GB for the 9B (8 GB with the 4B) |
| Disk | ~18 GB | ~10 GB (app + chat model) | ~10 GB (app + chat model) |
| Extras | — | — | zenity or kdialog (file dialogs); ALSA-compatible sound (PulseAudio / PipeWire) |
Optional voice / document / image packs need extra disk (see the table above).
Differences on Windows and Linux in this release: image questions use Qwen3-VL 2B; scanned / image-based PDFs are read, but PDFs whose pages are pure digital text are not rendered yet (attach a screenshot of the page instead — on the Mac every PDF works); audio attachments are WAV files (the Mac also takes MP3, M4A, AIFF, FLAC, CAF); drag-and-drop works on the Mac and Windows (on Linux use the attach button).
Verified on a Windows 11 laptop (Intel Core i7-13800H, 32 GB), a Debian 13 desktop (GNOME on X11), and Apple-silicon Macs.
✨ Using emu
- Chat with streaming Markdown (tables, code blocks with Copy, lists), saved chats, Copy / Regenerate, and Thinking for harder questions.
- 🎤 Dictate — Auto or 30 languages; Vietnamese numbers and dates written as digits.
- 🔊 Hear replies — Pocket TTS (English), Qwen3-TTS (multilingual), gwen-tts and VieNeu-TTS (Vietnamese), with voice cloning from a short recording. Every clip is saved (Clips) to replay or export.
- 📄 Read documents — PDFs and pictures of text into the chat, tables kept; the Documents library keeps them all.
- 🖼️ Ask about images — attach a picture and ask; Qwen3-VL answers.
- 100% private & offline — chats are plain files on your computer
(
~/Library/Application Support/emu,%LOCALAPPDATA%\emu,~/.local/share/emu); nothing leaves your device.
English and Vietnamese are first-class across chat, speech and voice; the chat models natively handle dozens more languages.
🕰️ Previous edition: emu 2.0 (Windows and Linux)
emu 2.0 remains available. It is the choice if you want NVIDIA CUDA GPU acceleration on Windows or Linux, or its larger model menu (Gemma 4 12B and E4B, Qwen3.6 27B, Qwen 3.5 4B / 9B). It comes as an installer in a CPU and a GPU (CUDA) edition.
| Your computer | Edition | Installer | Size |
|---|---|---|---|
| Windows + NVIDIA GPU (GTX 16-series / RTX 20-series or newer) | GPU (CUDA) | emu-2.0.0-windows-cuda-setup.exe | ~758 MB |
| Windows, no NVIDIA GPU | CPU | emu-2.0.0-windows-cpu-setup.exe | ~62 MB |
| Linux + NVIDIA GPU (GTX 16-series / RTX 20-series or newer) | GPU (CUDA) | emu-2.0.0-linux-cuda.run | ~866 MB |
| Linux, no NVIDIA GPU | CPU | emu-2.0.0-linux-cpu.run | ~119 MB |
emu 2.0 model packs (downloaded by emu 2.0's model picker):
gemma-4-12b-it-UD-Q2_K_XL.nbq, gemma-4-12b-it-qat-q4_0.nbq,
gemma-4-e4b-it-q4_k_m.nbq, Qwen3.6-27B-int4-AutoRound.nbq,
qwen3.5-4b-q4_k_m.nbq, qwen3.5-9b-q4_k_m.nbq.
emu 2.0 requirements: Windows 10/11 (64-bit) or Linux x86-64; a 64-bit CPU with
AVX2; 16 GB RAM (32 GB recommended); 8 GB free disk plus models.
CUDA editions: an NVIDIA GPU with Compute Capability 7.5+ (8.6+
recommended; 6–8 GB VRAM, 12 GB+ for the largest model) and a driver supporting
CUDA 13 (release R595+). The GPU runtime is bundled — no CUDA Toolkit,
cuDNN or developer tools needed; run nvidia-smi and check the CUDA Version is
13.x. If the GPU or driver is unsupported, emu runs on the CPU.
Install: Windows — double-click the .exe (code-signed by CLOUDKITES PTY
LTD), accept the licence and finish. Linux —
chmod +x emu-2.0.0-linux-*.run && ./emu-2.0.0-linux-*.run, then follow the
installer. On first run, open the model picker and let emu download a chat model
suited to your hardware.
📚 Documentation
Read these for full details (also on the Files tab):
- Download & Installation Guide (PDF) — choosing an edition, downloading, installing, and getting started.
- User Guide (PDF) — a complete, step-by-step guide to every feature on Mac, Windows and Linux, plus emu 2.0.
- Product Information (PDF) — capabilities, AI models and their licences, system requirements, privacy, and limitations.
📜 License
emu is proprietary software distributed free of charge © 2026 CloudKites Pty Ltd (CloudKites AI Lab), ABN 91 633 195 874. It is provided in compiled (binary) form for macOS (Apple silicon), Windows (x64) and Linux (x86-64).
- ✅ Free to use — personal or commercial, on any number of devices.
- ✅ Free to redistribute — share the complete, unmodified official package (DMG, zip, tar.gz or installer), free of charge, with all notices intact.
- ❌ No selling / no paid bundling, no modification, no reverse-engineering.
- The source code of emu and its Numbat ML toolkit is confidential and is not distributed.
"emu", "Numbat" and "CloudKites" are trademarks of CloudKites Pty Ltd.
Model licences. Qwen3.8 27B, Qwen3.5 9B / 4B, Qwen3-ASR, Qwen3-TTS and Qwen3-VL
© Alibaba Cloud (Apache-2.0); the Qwen3.8 27B 4-bit GGUF quantization by Unsloth AI
(Apache-2.0); the Qwen3.8-27B-DFlash2 draft model © z-lab / Inco AI (Apache-2.0);
Pocket TTS © Kyutai (CC-BY-4.0); gwen-tts © G-Group AI Lab (MIT); VieNeu-TTS ©
pnnbao-ump with MOSS-TTS-Nano © the OpenMOSS Team (Apache-2.0); PP-OCRv6 ©
PaddlePaddle (Apache-2.0); Unlimited-OCR (MIT). emu 2.0 also uses Google DeepMind's
Gemma models (Gemma Terms of Use). Weights are repacked into emu's single-file
.nbq format, unmodified apart from quantization.
License documents (all on the Files tab):
- End-User License Agreement —
LICENSE(plain text) — the full terms, included in every package. - End-User License Agreement —
LICENSE.pdf(formatted PDF) — the same terms. - Third-party attributions —
THIRD_PARTY_LICENSES.md— licences for the AI models, typefaces and libraries, each used under its own terms.
⚠️ AI output notice
emu uses generative AI models. Output may be inaccurate, incomplete, or fabricated even when it appears confident, and is not a substitute for professional advice — verify important information against trusted sources.
💬 Support & contact
Questions or issues? We're happy to help.
- Email: contact@cloudkites.com
- Website: https://cloudkites.com/contact