emu is a local-first, multimodal AI desktop application that runs
state-of-the-art AI models entirely on your own computer. Chat with text,
images and audio; dictate by voice; have replies read aloud; and extract text
from documents — 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 installer for free
(no selling, no paid bundling). See License below.
The installers are code-signed and self-contained — the GPU build bundles the
NVIDIA CUDA/cuDNN runtime, so there is nothing else to install. On first run, emu
offers to download the model bundle(s) you choose into your local models folder.
The Linux installers are self-contained (the Qt runtime, all required plugins, and
— for the GPU build — the CUDA/cuDNN runtime are bundled). They are distributed
over HTTPS and are not code-signed. After downloading, mark the installer
executable and run it:
bash
1chmod +x emu-1.0.0-linux-cuda.run # or emu-1.0.0-linux-cpu.run2./emu-1.0.0-linux-cuda.run
Launching emu after install: the installer adds emu to your application
menu — just click it. Or run the installed binary directly:
"<install-folder>/emu" # e.g. ~/emu/emu — runs from anywhere, no wrapper needed
Requires a glibc-based 64-bit Linux (glibc ≥ 2.38) with a working X11/XCB display.
Microphone / speech-to-text: emu records using your system's audio recorder, so
one of these must be installed (on a standard PipeWire desktop the first is already
present):
Getting a model: open the model menu at the top and pick one to download, or
just type a prompt and press Enter — if no model is installed yet, emu downloads
a recommended one automatically and sends your message once it's ready.
Multimodal chat — text, plus attach an image or audio clip.
🎤 Speech-to-text — dictate prompts on-device; auto-detect or choose from 20 languages.
🔊 Text-to-speech — read replies aloud with four engines, including
multilingual output and on-device Vietnamese voice cloning.
📄 Document OCR — extract text and tables from images and PDFs.
Multilingual — chat models natively handle dozens of languages; English
and Vietnamese are first-class across chat, speech and voice.
100% private & offline — every model runs on your CPU or NVIDIA GPU; nothing
leaves your device. No account, no telemetry.
Persistent chats, Markdown rendering, model picker, and resource-aware loading.
🧠 Models (in this repository)
Each model is a single self-contained .nbq bundle (weights + tokenizer +
helpers in one file), auto-downloaded by emu on first use. You only need the
model(s) for the features you want.
File
Capability
Base model · Provider · License
gemma-4-12b-it-UD-Q2_K_XL.nbq
Chat (text) — compact 2-bit, ~5 GB
Gemma 4 12B · Google DeepMind · Gemma Terms
gemma-4-12b-it-qat-q4_0.nbq
Chat (text), higher accuracy — 4-bit, ~7.5 GB
Gemma 4 12B · Google DeepMind · Gemma Terms
gemma-4-e4b-it-q4_k_m.nbq
Chat + vision + audio (multimodal)
Gemma 4 E4B · Google DeepMind · Gemma Terms
qwen3.5-4b-q4_k_m.nbq
Chat (text), light & fast
Qwen 3.5 4B · Alibaba Qwen · Apache-2.0
qwen3.5-9b-q4_k_m.nbq
Chat (text), stronger
Qwen 3.5 9B · Alibaba Qwen · Apache-2.0
qwen3-asr-0.6b.nbq
Speech-to-text (20 languages)
Qwen3-ASR · Alibaba Qwen · Apache-2.0
qwen3-tts-0.6b.nbq
Text-to-speech (multilingual, preset voices)
Qwen3-TTS · Alibaba Qwen · Apache-2.0
gwen-tts-vn-0.6b.nbq
Text-to-speech — Vietnamese voice cloning
gwen-tts · G-Group AI Lab · MIT
pocket-tts-en.nbq
Text-to-speech — fast English streaming
Pocket TTS · Kyutai · CC-BY-4.0
vieneu-tts.nbq
Text-to-speech — Vietnamese / English
VieNeu-TTS · pnnbao-ump · OpenMOSS · Apache-2.0
ppocr-v6.nbq
Document OCR (detection + recognition)
PP-OCRv6 · PaddlePaddle · Apache-2.0
Bundled AI models remain under their own licenses — see
THIRD_PARTY_LICENSES.md
and each model's own terms. Use of the Gemma models is additionally subject to the
Gemma Terms of Use and Prohibited Use Policy.
💻 Which model for my computer?
Your PC
Recommended chat model
Notes
NVIDIA GPU, 12 GB+ VRAM
Gemma 4 12B (4-bit); Qwen 3.5 9B
Highest quality, fast
NVIDIA GPU, ~8 GB VRAM
Gemma 4 12B (2-bit); Qwen 3.5 4B
Fits an 8 GB card
CPU only (AVX2, 16 GB+ RAM)
Qwen 3.5 4B; Gemma 4 E4B
Quick without a GPU
Image / audio understanding
Gemma 4 E4B (multimodal)
CPU or GPU
The GPU (CUDA) build gives real-time text-to-speech and OCR in seconds per
page. The CPU build runs every feature (OCR and the largest models are
slower). emu's Auto device setting uses the GPU when available and falls back
to the CPU otherwise.
✅ Free to use — personal or commercial, on any number of devices.
✅ Free to redistribute — you may share the complete, unmodified official
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.
Full terms: LICENSE.
"emu", "Numbat" and "CloudKites" are trademarks of CloudKites Pty Ltd.
⚠️ 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.
