What's New in v2.2.0
GGUF Loader v2.2.0 (August 11, 2026) upgrades Agent Mode to a proper LangGraph state machine with resumable sessions, adds a one-click GPU installer right in the app, ships separate GPU and CPU builds for Windows, and makes Find Paragraph dramatically faster on big folders.
🖥️ GPU & CPU Builds for Windows
Windows now ships two standalone executables, so you only download what your machine needs:
GGUFLoader_v2.2.0_CPU.exe(~69 MB) — runs on any PC, no GPU required. Great for laptops and office machines.GGUFLoader_v2.2.0_GPU.exe(~854 MB) — bundles the NVIDIA CUDA runtime for GPU-accelerated inference. Pick this if you have a CUDA-capable NVIDIA card.
The GPU build was previously the only option — the CPU-only build is 12× smaller and installs in seconds.
⚡ Much Faster Find Paragraph Search
Searching a folder with Find Paragraph is now up to ~40× faster on large trees. The search walk now:
- skips dependency and build directories (
.venv,node_modules,.git, …) during the walk instead of scanning every file, and - avoids repeated regex work when normalizing paragraph text for deduplication.
On a typical project folder the search went from ~0.8 s to ~0.02 s with identical results.
🚀 Hardened Launchers for Linux & macOS
launch.sh (and launch.bat on Windows) now check everything up front and tell you exactly what to do:
- Python 3.10+ — clear error if missing or too old
- Broken venv detection — catches the Ubuntu/Debian
python3-venvstub and falls back tovirtualenv - No compiler needed — Linux installs automatically pull the prebuilt CPU wheel from the llama.cpp project's index, so you no longer need gcc/cmake to run from source
- Network preflight — checks PyPI is reachable before installing, and saves a full log to
/tmp/gguf_pip_install.logif something goes wrong
🤖 LangGraph-Powered Agent Mode
Agent Mode is now built on LangGraph — a proper state machine (START → agent → tools → agent → END) instead of an ad-hoc loop. That brings real engineering benefits:
- Resumable conversations: every workspace's thread is checkpointed to SQLite, so closing the app and reopening it resumes the exact same conversation — even across restarts
- Approval before sensitive actions: shell commands and git writes suspend the graph and ask you to approve before anything executes
- Cleaner runs: step budgets, malformed-JSON repair, corrective retries on failed tool calls, and cooperative cancel — stop a run at any time and it stops cleanly
- Streaming answers: the final summary streams token-by-token into the chat
🖱️ One-Click GPU Install
No more downloading wheels or running batch scripts by hand. The sidebar now has an ⬇ Install GPU Support button that:
- detects whether CUDA support is already installed
- downloads and installs the CUDA-enabled llama.cpp build in the background with a live status line
- shows ✅ GPU Support Installed when done — just restart the app and pick GPU mode
If you try to load a model with GPU acceleration before installing, the app tells you exactly where to click.
📚 Documentation
- New AGENTS.md — how AI agents should work on this codebase (build, test, addons, release)
- Rewritten QUICK_REFERENCE.md and a new addon development guide (
ggufloader/addons/README.md) with a working example - The full
docs/archive (user guide, FAQ, addon API, build instructions) was restored from git history and linked from the README - Build scripts now auto-name artifacts
GGUFLoader_v<version>_<GPU|CPU>.exeandGGUFLoader_v<version>_linux_x86_64_CPUfrom the single version source of truth
Download v2.2.0
- Windows — GGUFLoader_v2.2.0_CPU.exe (~69 MB, works everywhere)
- Windows — GGUFLoader_v2.2.0_GPU.exe (~854 MB, NVIDIA CUDA)
- Linux — GGUFLoader_v2.2.0_linux_x86_64_CPU (standalone binary, chmod +x and run)
- All releases on GitHub
💡 Which Windows build? Not sure if you have CUDA? Use the CPU build — it runs on any machine. Choose the GPU build only if you have an NVIDIA graphics card and want maximum speed.