Cloud images (AutoDL / UCloud)

For people who do not want to install Docker or drivers: rent a 24 GB GPU instance, pick the Mynah image, run one command, open an SSH tunnel, talk to Luna in the browser.

Positioning: single-node trial build

The image puts every component on one GPU in one instance, for trials, development and demos. Mynah itself is an enterprise-grade, production-oriented architecture (see Architecture): containerised deployment, worker pools with concurrency quotas, per-channel publishing, knowledge base and instinct layer all live in the same codebase. For production follow the Deploy guide and run it multi-instance; the configuration and data from the image carry over.

Status

The recipe, bootstrap script and mynahctl live in deploy/cloud/ and the whole chain has been run end to end in an AutoDL-like CUDA container. The platform images are being published; links land here when they are live. Until then the manual steps below take about fifteen minutes on any CUDA 12 / Ubuntu 22.04 base image.

What is in the image

Component	Notes
FlashHead (Lite weights)	default avatar engine, half-body canvas, about 6.5 GB VRAM
Luna / Nova built-in presets	with baked idle breathing and shy / laugh / shake / wink clips
SenseVoice + Smart Turn	speech recognition + semantic end-of-turn
EdgeTTS	default voice; switch to Qwen3-TTS for cloning
qwen2.5:7b via ollama	the brain. Pulled on first start on AutoDL (data disk); bundled on UCloud
PostgreSQL 16 + pgvector	console, knowledge base, channel publishing
cored + console	realtime core and admin, all native, no Docker

Usage

~/mynah/mynahctl start        # starts everything; prints the first console password at the end

Browsers cannot reach these instances directly (AutoDL has no public UDP), so open a tunnel from your own machine that forwards the visitor page, the console and the WebRTC TCP port together:

ssh -N -L 8443:127.0.0.1:8443 -L 9443:127.0.0.1:9443 -L 8444:127.0.0.1:8444 -p <port> root@<instance>

Open https://127.0.0.1:8443/ (self-signed certificate, click through), allow the microphone, talk. Console at https://127.0.0.1:9443/.

Why three ports

WebRTC prefers UDP; an SSH tunnel only carries TCP. The image starts cored with --ice-tcp :8444 --ice-nat-ip 127.0.0.1, so the browser receives a passive TCP host candidate at 127.0.0.1:8444 and media flows through the tunnel. On a UCloud instance with a public IP, set ICE_NAT_IP in ~/mynah/env to that IP, open the three TCP ports, and skip the tunnel.

Everyday: mynahctl status | logs <service> | restart | password; ports, GPU and model live in ~/mynah/env.

GPU

24 GB runs the default set (FlashHead ≈ 6.5 GB + 7B model ≈ 5.5 GB + ASR ≈ 1.4 GB). 16 GB works. On 12 GB set LLM_MODE=none and point the console at any OpenAI-compatible cloud model.

Build it yourself (what the image does for you)

On a fresh CUDA 12 / Ubuntu 22.04 instance:

git clone https://github.com/honwee/mynah.git ~/mynah/src
bash ~/mynah/src/deploy/cloud/bootstrap.sh                    # add --llm bundled on a big system disk (UCloud)
~/mynah/mynahctl start

bootstrap.sh installs system packages; downloads the cored release binary and the Luna/Nova preset clips; creates one conda env for every worker; installs PostgreSQL+pgvector from conda-forge and initialises it; clones FlashHead and downloads only the Lite weights (~8.4 GB, via hf-mirror); fetches SenseVoice from ModelScope; installs ollama; writes a self-signed certificate and ~/mynah/env. Downloads go through China mirrors by default and the script resumes when re-run.

Using the instance as the GPU for a laptop

You can also run only the avatar worker there and keep cored and the console on your own machine: forward 9421 (the FlashHead worker's gRPC port) and set AVATAR_ADDR=127.0.0.1:9421 locally. See Quickstart, path C.