Kimodo.cpp One-Click Windows Installer - Local AI Text-to-Motion 3D Animation
Type "a person waving" and get a 3D character animation you can drop straight into Blender or Unreal Engine 5 — generated locally on your own machine, no cloud, no subscription. Kimodo is NVIDIA's open text-to-motion model; this is my one-click Windows package for Kimodo.cpp, built so it actually installs on a normal Windows PC.
Why this exists: the original project had no real Windows install path — you were expected to build it from source with a full C++ toolchain. I forked it, added proper Windows support (the fork is free and open source at TheLocalLab/kimodo.cpp-windows), and packaged it so there is nothing to build.
Nothing to install first. No compiler, no Visual Studio, no CMake, no Vulkan SDK, no Go — not even Python. Every executable is prebuilt, including the web UI server, and setup bootstraps its own private Python into the package folder. Unzip, double-click Setup.bat, and it fetches the model weights on first run.
Generate from a local browser studio or a double-click command-line mode, preview the animated skeleton, and export a .glb to retarget onto the UE5 Mannequin or a Blender rig. Runs on an NVIDIA GPU with 8 GB of VRAM via Vulkan, or on CPU if you don't mind waiting.
What you get
The one-click supporter package for Kimodo.cpp on Windows:
Prebuilt Kimodo binaries and the GGML runtime libraries
The one-click installer (Setup.bat) plus three launchers — generate, web UI and demo
A model-weight downloader — resumable and SHA-256 verified
A condensed licence reference for the modelsModel files (downloaded automatically on first run — about 16.3 GB):
Llama-3 text encoder bundle (35 files) — ~15.2 GB — required
SOMA RP v1.1 motion model (30-joint rig) — ~1.13 GB — the defaultEverything stays inside the package folder. Nothing is installed system-wide, and uninstalling is just deleting the folder.
Current limits, upfront: the preview shows the animated skeleton, not a textured mesh. The SOMA 30-joint rig is the shipped default; the larger 77-joint expansion isn't implemented yet. SMPL-X is not included — its licence forbids redistribution.
System requirements
OS: Windows 11, 64-bit. Only tested on Windows 11 — Windows 10 may work but is untested.
GPU: an NVIDIA GPU with 8 GB VRAM runs it well via Vulkan. It also runs on CPU, just slower.
Disk: about 20 GB free — roughly 16.3 GB of model weights, around 500 MB for the private Python, plus working space. (An earlier "about 8 GB" figure going around is wrong — plan for 16–20 GB.)
Prerequisites: none. No coding required and nothing to install beforehand.
Getting started
1. Unzip the package into its own dedicated folder, e.g. C:\Kimodo\ — avoid Desktop, Documents, and cloud-synced folders.
2. Double-click Setup.bat.
3. Wait for it to finish. You're looking for 35/35 (text bundle verified) and valid (model smoke test passed).
4. Double-click Launch-Kimodo-UI.bat to open the browser studio at http://127.0.0.1:8094.
5. Start small to test: 60 frames, 20–30 steps. Generation takes roughly 5–20 minutes depending on your settings and hardware.
Troubleshooting
The animation doesn't play: it doesn't autoplay — click the entry under Past animations in the studio.
The viewport shows only a grid: hard-refresh the page with Ctrl+F5.
"ggml_vulkan: Found N devices" in the console: that's normal, not an error.
Setup stopped before 35/35: the weight downloader is resumable and verifies each file — run Setup.bat again and it picks up where it left off.
Install fails or behaves oddly: check the path. A cloud-synced folder or a deep path with spaces is the most common cause — move to C:\Kimodo\ and re-run.
Licence & credit
Kimodo is by NVIDIA; Kimodo.cpp upstream is localai-org/kimodo.cpp. The Windows fork is mine and free on GitHub (TheLocalLab/kimodo.cpp-windows), along with the full manual build path and an Unreal retargeting guide. This product is the prebuilt, one-click packaging of that work.
Licences:
Code: Apache-2.0 (the fork and the upstream project).
SOMA / G1 motion checkpoints: NVIDIA Open Model License — commercial use permitted.
Text encoder bundle: Meta Llama 3 terms apply.
SMPL-X: internal R&D licence only — not included; don't redistribute it.