Repetere AI 1.0 Sovereign Workstation

● User Manual · v1.0 Sovereign Workstation

Master your sovereign machine.

Everything from first install and custom weight swapping to hybrid VRAM offload and the v1.0-exclusive DeepThink Autonomous Lab — complete with real-time operator walkthrough videos.

Chapter 01 · Install & First Launch

Video 01 Included

Install the <90 MB native host & confirm hardware

  1. Download the installer for Windows 10/11 (64-bit) or Linux (AppImage / .deb) from Free Beta (v0.9) or Commercial License (v1.0). Note: Free Beta v0.9 includes the five core studios; the DeepThink Autonomous Lab (Chapter 06) is exclusive to v1.0.
  2. Verify the offline checksum (SHA-256 printed on your license receipt) — no online activator phones home.
  3. Run the installer. It places one native binary (~90 MB), a model vault folder, and per-studio config files. No services, no startup daemons.
  4. Hardware Confirmation: On first launch, Repetere AI detects your GPU(s) (supporting single and multi-GPU setups). Select your target GPU and click Confirm Hardware.
  5. Health & Concierge Navigation: Verify the bottom status strip shows Healthy subsystems, then use the workspace Concierge bar (e.g., type chat) to jump directly into any studio tab.
vault/ models/ configs/ exports/
host: repetere-ai (< 90 MB) · gpu: Confirmed → Subsystems: Healthy ✓
Walkthrough Video 01 · Your First Launch & Setup
Watch on YouTube ↗

Chapter 02 · Model Vault & Bring-Your-Own Weights

100% Model-Agnostic

Swap any compatible local checkpoint & tune per-model overrides

Repetere AI is completely model-agnostic — you are never locked into bundled weights. Drop any compatible local GGUF, SafeTensors, ONNX, or diffusion checkpoint directly into your models/ folder, open Settings → Repetere Models Manager, inspect All Installed Weights, and click Set Active for any studio tab. Each tab tracks readiness independently (switching from Red to Green Ready) so studios never collide in VRAM.

Bring Your Own Checkpoints

GGUF · SafeTensors · ONNX · Diffusion

Load custom fine-tunes, uncensored quants, or domain-specific weights from any local source and hot-swap them per studio anytime.

Per-Model Custom Overrides

GPU Layers · Context · KV Quant

Override baseline defaults on any individual weight file — control GPU layer offload (-ngl), context window size, KV cache quantization, and tensor splits.

Suggested Starting Tiers

Qwen 2.5 4B–14B · SD/FLUX · Audio

Want a baseline? Start with Qwen 2.5 Coder (8B GPU-Only / 14B Hybrid), ~8 GB SD/FLUX checkpoints, or Auto VRAM Guard audio weights (see storage sizing).

Chapter 03 · Core Studios (v0.9 & v1.0)

Videos 02, 03 & 04 Included

Run each studio — step-by-step & live walkthroughs

1. Coding IDE Studio

Project Scanner · Vibe Coder · DeepThink Triage
  • Local Project Workspace: Click Open Folder to scan and load your local source tree, edit files directly in the native window, and click Compile Project in-place.
  • Local Coding Agent (Vibe Coder): Generate complete applications from natural language prompts completely offline with zero cloud telemetry.
  • DeepThink Triage Bridge (v1.0): Load compiler or runtime logs from DeepThink runs, toggle Force Vectoring (compresses file context for speed) and Enforce Local Code Fences (emits full replacement blocks), and click Diagnose & Generate Fix Directive to produce a targeted repair prompt for DeepThink.
Walkthrough Video 02 · Coding Tab & Triage Watch on YouTube ↗

2. Chat & Story Bible (Creative Writing) Studios

Prompt Formatters · Lore RAG · Live Editor
  • Offline Chat Assistant: Full-context local conversational workspace for brainstorming, architecture planning, and coding discussions with zero token fees.
  • Built-in Studio Prompt Formatters: Use Format Song for Music Studio to turn plain descriptions into structured lyrics and genre tags ready for one-click copy into Vocal Music, or use DeepThink Prompt Writer (with the Targeted Fix toggle) to craft multi-step build/iteration directives.
  • Story Bible & Lore RAG: Automatically injects character sheets, world-building files, and project lore into the system prompt so chapters stay strictly consistent with your canon.
  • Real-Time Document Editor: Draft stories, run inline AI revisions (e.g., rewriting scenes or introducing characters in real time), use Find & Replace, Undo/Redo, and auto-save local revisions.
Walkthrough Video 03 · Chat & Creative Writing Tabs Watch on YouTube ↗

3. Vocal Music & Diffusion (Image) Studios

Voice Cloning · Photoshop-Lite · Local Inpainting
  • Vocal Music Synthesis: Paste formatted lyrics and style tags, name your track, and render full offline audio (~30 seconds generation per segment) with automatic built-in waveform playback.
  • Custom Voice Clone Profiles: Record and save your own voice profile, select it in the studio, fine-tune vocal parameters, and generate original songs sung in your cloned voice.
  • Diffusion Text-to-Image: Generate high-resolution local images (~8 GB VRAM footprint on 12 GB GPUs) with negative prompt controls (what not to do) and automatic low-VRAM fallback support.
  • Photoshop-Lite Editor & Local Inpainting: Apply real-time filters, gradient mixes, brightness, and saturation adjustments without regenerating — or brush/highlight any region to run localized prompt-driven Inpainting replacements.
Walkthrough Video 04 · Music Studio & Image Studio Watch on YouTube ↗

Chapter 04 · Dual Engine & Offload Modes

Hot-swap CUDA ↔ Vulkan & toggle GPU-Only vs. Hybrid Offload

Switch between CUDA (NVIDIA RTX) and Vulkan (Radeon / Arc / integrated) directly from the studio bar without restarting the host. When balancing custom weights against context window size, toggle between two execution modes:

GPU-Only Mode

Loads 100% of model layers into VRAM for peak token generation speed and massive context windows. Ideal for 7B/8B Chat and Story Bible sessions.

Hybrid Offload Mode

Reloads the instance server and splits layers across VRAM and system RAM — essential when running heavy 14B+ high-quant models on 12 GB VRAM cards without sacrificing context window size.

Recommended hardware pairings are listed under System Requirements.

Chapter 05 · Troubleshoot

Fix it offline

Out of VRAM on 14B+ models or Audio/Diffusion?

Switch from GPU-Only to Hybrid Mode, drop one quantization/model tier in Models Manager, or close an idle studio. The host never pages models to the cloud.

Engine won't bind?

Update GPU drivers, then re-confirm your active GPU from first-boot hardware detection or Help → Re-detect GPU.

Slow diffusion or UI freeze warning?

Pin generation to CUDA on RTX (~8 GB VRAM footprint), acknowledge the memory guard prompt before generating, and use real-time filters/inpainting instead of full re-renders.

Chapter 06 · DeepThink Autonomous Lab

Exclusive to v1.0 · Working Active Beta
Video 05 Included

Autonomous reasoning, Python AST splicing & live telemetry loops

Exclusive to Repetere AI 1.0 (not included in the v0.9 Free Beta), the DeepThink Autonomous Lab is a working, shipping autonomous build engine currently in active beta on our roadmap. Engineered separately from the five core studios, it executes recursive multi-step reasoning, builds complete C++, Python 3, or HTML5 applications offline, and scales from 8 GB VRAM Windows handhelds up to multi-GPU workstations.

1. Workspace Setup & Automatic Loop Snapshots:

Set your target working folder, choose any compatible local coder model override in Models Manager (e.g., Qwen 2.5 Coder 7B or 14B in Hybrid mode), and select your target stack. Every click of Begin Loop automatically snapshots your files so you can Revert to Last Loop at any time (keeping a manual backup of untouched source is also recommended).

2. Python Abstract Syntax Tree (AST) Splicing:

Instead of dumping and rewriting entire large files on every pass, DeepThink parses code into an Abstract Syntax Tree, splices targeted changes cleanly, and verifies execution before advancing.

3. Live Telemetry Output & Continuous Roadmap Upgrades:

A real-time telemetry window reports every phase and step across a typical 5–8 minute local autonomous pass. Once validated, inspect the build in the integrated HTML preview window or click Launch in Default Browser, iterating with follow-up loops or Coding Tab triage directives as new roadmap improvements roll out.

Walkthrough Video 05 · DeepThink Autonomous Lab (Real-Time Build)
Watch on YouTube ↗