Get up and running

Choose the stable release or Linux beta

The download section separates stable 1.0.24 from the 2.0.0-dev.6 Linux beta. Check the beta manifest for its exact package hash, requirements and test scope before installing.

• Debian/Ubuntu amd64 beta — download Saient_2.0.0-dev.6_amd64.deb and install it with your package installer or sudo apt install ./Saient_2.0.0-dev.6_amd64.deb. Back up your projects and app data before trying a beta.
• Stable Linux — the 1.0.24 Debian package and AppImage remain available. For the AppImage, allow execution in file properties or run chmod +x Saient_1.0.24_amd64.AppImage.
• Stable Windows x64 — Saient_1.0.24_x64-setup.exe remains available. It is not Authenticode-signed, so Windows may show an unknown publisher. This is not a Windows build of dev.6.

The dev.6 release is a Linux beta with known coding reliability limits. Native Windows, new-machine setup and broader hardware coverage have not been established by its Linux checks.

First run & setup

The first time you open Saient, a setup wizard walks you through getting ready:

• Fast setup — installs the essentials and bundled AI engine. You still need to choose or download a compatible local model.
• Full setup — also installs the optional Python tools (in a managed virtual environment) for image generation, the vision analyzer, text-to-speech and LoRA training.

You can re-run setup or install extras later from inside the app.

Loading models

Chat and Agent use GGUF language models. Select an installed .gguf file in the model picker. Model architecture, quantisation, context size and memory must all fit the engine and your machine; a filename alone does not establish compatibility.

Creative tools use separate diffusion models. Image Gen and the SDXL LoRA trainer scan supported image models. Merge uses compatible single-file diffusion checkpoints, not GGUF chat models or Diffusers folders.

GPU & model memory

See the beta manifest requirements for the packaged engine and supported hardware scope. NVIDIA acceleration needs a compatible driver; nvidia-smi checks the driver, but does not prove that every bundled engine or creative model supports your GPU.

The title bar shows resident models across screens and provides Unload / Free VRAM. Stopping image generation keeps the loaded model available for another run; stopping its loading stage may terminate the worker. Explicitly unload when you want to release that memory.

The studios in the Linux beta

Switch screens from the left rail. Ctrl+1 through Ctrl+9 follow Chat, Agent, Workflows, Image, Assets, Video, TTS, LoRA and Merge.

• Chat — local model chat with streaming, Stop, optional Saient binding and Dual Agent.
• Agent — project files, terminal, Planner, memory and checkpoints. Read the agent guide before granting execution permissions.
• Workflows — Saient's native node editor, with text, templates, image generation, imported images, masks and previews. Save/load JSON, undo/redo and run to a selected node are available; ComfyUI is not required.
• Image Gen — supported diffusion models, LoRAs and sampling controls. Inpainting edits a painted region using supported SD/SDXL models; SD3/3.5 inpainting is not implemented.
• Assets and Video — manage local outputs and use supported local video workers. Availability and memory needs depend on the model.
• Text-to-Speech — local speech generation and WAV saving.
• LoRA & Merge — SDXL adapter training and diffusion-checkpoint merging. These require the optional Python tools.

Workflow JSON supports Saient's own node types. It does not import ComfyUI graphs or custom nodes.

Free local inference

Saient is free to download. There is no Saient account, subscription, trial timer or licence-key screen.

Permissions and the launch password

Planner Write mode, terminal confirmations and /yolo, and Long Project permissions are separate controls. Enabling one does not grant the others. Terminal file tools enforce their workspace boundaries, but authorised shell commands run with your user account's access. Selecting a workspace is not an operating-system sandbox.

Long Project saves its goal, permission settings, budget and configured verification command. General shell execution is separate from file writes; the exact configured verifier has its own explicit authorization. Review command effects before granting it. A passing check is evidence for that check, not proof of every requirement.

The optional launch password deters casual app access. It does not encrypt files on disk or confine commands.

Troubleshooting

A model won’t load or the app feels stuck

Check the visible error, model compatibility and available RAM/VRAM. Cold loading from a hard disk can take substantially longer than a warm run. Use Stop to cancel active work, and Unload / Free VRAM to release a resident model before retrying.

Image gen / vision / TTS aren’t available

These need the optional Python tools. Re-run Full setup to install them.

The GPU isn’t being used

Confirm the driver works with nvidia-smi, then compare your hardware with the beta manifest requirements. CPU language inference and GPU creative workloads have different requirements; do not treat a working chat model as a test of image generation.

I see a trial or licence prompt

You have an obsolete installer. Uninstall that copy and use the current free build from the download section. The current build does not require a Saient key.