What's New in v2.3.0
GGUF Loader v2.3.0 (September 11, 2026) is a testing release that takes the agentic experience further with a strictly plan-driven agent and a developer-style inline process UI. It is temporarily pinned to a single model (Gemma 4 12B Instruct Q4_K_M) to tune the agent's performance before expanding model support.
⚠️ Testing Release: This version is locked to one model while we validate the new plan-driven agent architecture. Universal model support returns in the next release.
🎯 Single-Model Focus (Temporary)
This testing release is optimized for Gemma 4 12B Instruct Q4_K_M. Multi-model detection was temporarily removed so this one model just works. Universal loading returns in the next release.
📋 Strictly Plan-Driven Agent
Every agent turn now goes through a planner node:
- Tool-free questions — answered directly without planning
- Tasks — get a step-by-step plan executed with sandboxed tools
- The reactive ReAct fallback was removed for more predictable behavior
🖥️ Developer-Style Process UI
Plan steps, tool calls, and results render inline in the chat above each answer. You can follow the entire agent process as it happens.
🧹 Reliable Final Answers
Stray tool-call JSON envelopes and stale status text (like "Planning...") are scrubbed from the live stream and final replies.
📥 Auto-Load at Startup
The app scans its models/ folder (plus the last-used model folder) and loads the pinned GGUF automatically. If the model is missing, the model chip in the header offers a Download action with live progress.
🔧 Pruned Tools
The default registry is the 10 workspace tools the agent actually uses — memory and meta tools were removed, and searches are hardened against huge generated folders.
🚀 Self-Setup Launchers
launch.bat and launch.sh now handle all dependency installation and offer browser, Electron, and production modes.
Full Capabilities
Agent Tools
| Tool | What it does |
|---|---|
list_directory / glob |
Explore folders and match file paths in the workspace |
read_file |
Read any file (MD/PDF/DOCX/TXT/code — text extracted automatically) |
search_files |
Find files and grep for content |
write_file / edit_file / move_file |
Create, edit, and move files |
run_command |
Run a shell command inside the workspace (approval-gated) |
run_python |
Execute Python source inside the workspace (approval-gated) |
git |
Git operations — read-only runs freely, writes require approval |
Advanced Search (Find Paragraph, no RAG)
- Single file — type a question and the model finds and ranks matching paragraphs
- Folder search — a planner decides which files to read, with live per-file progress
- Smart defaults — your last query, source, and settings are remembered between sessions
Interface Features
- Sessions left, chat center, tools right layout
- Streaming chat via WebSocket
- Model chip in header with live download progress
- Settings dialog (Model, Providers, Agent, Hardware, Appearance, Keyboard, Plugins)
- Dark/light theme with accent colors
- Command palette (Ctrl/Cmd+K)
- Electron standalone desktop app
Core Features
- Plan-driven agent (LangGraph) with Allow/Deny approval cards
- SQLite checkpointing — conversations survive restarts
- Real file reading (MD/PDF/DOCX/TXT/code)
- GPU acceleration via Settings → Hardware
- 100% local inference — no data leaves your machine
- Cross-platform (Windows 10/11, Linux, macOS)
Download v2.3.0
🪟 Windows Downloads
- GGUFLoader_v2.3.0_CPU.exe (~145 MB) — Works on any PC, no GPU required
- GGUFLoader_v2.3.0_GPU.exe (~930 MB) — NVIDIA CUDA for faster inference
🐧 Linux Download
- GGUFLoader_v2.3.0_linux_x86_64_CPU (~50 MB) — CPU-only, chmod +x and run
💡 Which Windows build? Use the CPU build if you're not sure — it runs on any machine and is 6× smaller. Choose the GPU build only if you have an NVIDIA graphics card for maximum speed.
First Launch
- Start the app — it auto-loads the pinned Gemma 4 12B Q4_K_M from the
models/folder in the background - No model on disk? The model chip in the header downloads it with live progress and loads it when finished
- Chat in the main window, or press Ctrl/Cmd + Shift + A for Agent Mode and choose a workspace folder
System Requirements
- OS: Windows 10/11, Linux, macOS (Intel & Apple Silicon)
- RAM: 32 GB minimum recommended
- Storage: ~8 GB free for the model file
- GPU: Optional — NVIDIA CUDA on Windows/Linux, Metal on macOS
What's Coming Next
🔮 Universal Model Loader returns — the next release removes the single-model restriction. You'll be able to run any GGUF model with the agent, with hardware-aware recommendations so you can pick a model that fits your PC's RAM/VRAM.