kimodo.cpp ports NVIDIA's text-to-motion model to GGML and runs it on CPU or Vulkan in 48 hours
GitHub·medium signal
localai-org/kimodo.cpp reached 375 stars and 29 forks in the two days after its 2026-08-22 creation. It runs Kimodo-SMPLX-RP-v1 from either a UTF-8 prompt or a precomputed LLM2Vec embedding, generating unconstrained SMPL-X22 local rotations and root translations, with GGUF loading, safetensors conversion, DDIM sampling, C/C++ APIs and CPU/Vulkan parity tests. The text encoder defaults to eight-layer Vulkan chunks with KIMODO_TEXT_LAYER_CHUNK=1..32 to tune VRAM; constraints, SOMA, G1, GLB export and quantized models are not implemented yet.