Saient is a local AI desktop app — chat, a real coding agent, image & video generation, a vision analyzer, text-to-speech, and LoRA training, running on your own hardware. Local inference needs no Saient account, licence key or hosted-inference key.
[>_<]
/|#|\
/ \
Task-focused local AI tools in a single native window. Language, image, video and speech use model-specific local runtimes.
Talk to any local GGUF model. Renders live HTML artifacts in a side pane — ask it to build a tool or a game and watch it run.
A real agent with a PTY terminal and a tool-use loop (read · ls · write · edit · bash). It plans, writes files, runs commands, and fixes its own errors — on your local model, in a sandboxed workspace.
SDXL text-to-image with LoRA support, schedulers, live progress, and an optional face-detail pass.
Text-to-video and image-to-video with Wan & CogVideoX — frame control, live progress, and an optional refine & upscale pass.
Ask questions about any image — describe it, read its text, count objects — with a local Moondream2 analyzer. Paste a screenshot and go.
Kokoro American and British voices with speed control and local 24 kHz WAV output.
Train SDXL LoRAs with a dataset cleaner. A local checkpoint merge worker exists; the v1.0.10 picker limitation is documented.
Powered by Quartz, Saient's native inference runtime — local model execution with no ONNX Runtime or cloud inference.
Step-by-step pages with real screenshots, exact model boundaries, hardware notes, proof status, downloads and source links.
Install, add a model, choose CUDA or CPU, tune settings and inspect a real image output.
Text-to-video, image-to-video, actual model routes, memory boundaries and MP4 proof.
GGUF setup, architecture support, RAM/VRAM sizing and a live local response check.
The direct boundary between host-model proposals and Saient's tools, state, memory and control loop.
A full evidence-first paper on durable memory, checkpoints, goal continuity, long-horizon agency and the failures behind the claims.
Dataset preparation, exact trainer scope, hardware unknowns and the current proof boundary.
Weighted and add-difference checkpoint merges, memory behavior and the confirmed v1.0.10 picker defect.
Kokoro voices, CPU execution, WAV output and current evidence status.
A direct comparison: focused studios and agent versus node graphs, custom nodes and wider model coverage.
Open the Agent screen, choose a project and local model, then give it a scoped goal. Saient can run a plan→act→verify loop within the permissions you grant.
[>_<] /|#|\ / \
Download the free desktop app, add the models you choose, and run supported AI workflows on your own hardware.
On code signing — we’re on it. The Windows installer is not yet Authenticode-signed, so Windows will warn about an “Unknown publisher” and SmartScreen may ask you to confirm. That is a certificate we are actively working to put in place, not a corner we cut. The release pipeline is already wired to sign automatically the moment one is available, so this notice disappears on its own. Until then every download is listed with its SHA-256 in the release manifest so you can check what you got.
Platform verification. Both packages contain the same terminal and runtime repairs. Version 1.0.24 passes native Windows runtime, frontend and Rust CI checks; an installed Windows desktop/GPU session has not been verified for this version. Linux desktop testing remains the more extensive of the two. Exact source revisions, checksums and verification limits are recorded in the release manifest.
Windows prerequisites. The bundled inference engines require the Microsoft Visual C++ v14 Redistributable (x64). Install it from Microsoft if it is missing. Python 3.10+ is also required for the Saient terminal. NVIDIA GPU acceleration requires a compatible NVIDIA driver. The CUDA runtime library is bundled; the CUDA toolkit is not required.
Thank you. To everyone who has downloaded Saient, reported something broken, or just followed along — genuinely, it helps, and it is the reason the rough edges get found at all.