Hinode v3 is out: set up once, run on any GPU

In v2 your setup lived on one machine. A bigger card meant moving the whole desktop to it. In v3 your setup and your GPU are two separate things.

One workspace, many machines. The workspace is a small Linux computer with a desktop and an address that does not change, at $0.30/hr. Install your tools there once and save a snapshot. Every machine you start copies it.

Pick the card the model needs. An L4 with 24 GB, an L40S with 48 GB or an RTX PRO 6000 with 96 GB, each with 32 to 256 GB of RAM. That makes eleven sizes, from $1.75 to $12.58/hr. If no GPU of that size is free, Hinode keeps trying for an hour, and billing starts only when the machine does.

Install ComfyUI in one click. ComfyUI, PyTorch, JupyterLab, Ollama, AI Toolkit, VS Code and more install from the workspace screen, together with what they need. You can keep working while they install.

One folder for every machine. Models, datasets and outputs live in ~/shared. A LoRA trained on one machine shows up on another as soon as it is saved. You download a checkpoint once.

Start a training run and go to bed. hinode run hands a command to a machine. Close the laptop. You get an email when the job ends, with the last lines of the log.

Idle machines pause themselves. After 15 quiet minutes the machine pauses and keeps its disk. A training run that is only downloading data still counts as work. Disks grow on their own, and you pay only for what is written: $0.15/GB-month.

Ask Claude to start the run. Hinode is in Claude's connector directory. ChatGPT, Claude Code and Codex can use the same account, and they can start machines and run jobs for you.

Under that: SSH works with one command, and VS Code works over it. The shared folder costs $0.11/GB-month. A monthly budget sends you an email and stops nothing. The Linux app now matches the Mac, and there is a light theme.