Llama 3 Kimodo Turbo

Built with Meta Llama 3.

Generate human motion from text on Apple Silicon. Complete 16.3 GB inference bundle: Kimodo-SOMA-RP-v1, normalization statistics, tokenizer and preconverted LLM2Vec/Llama 3 encoder. Use the Kimodo Turbo runtime; no local weight conversion is needed. The encoder streams BF16 layers with FP32 LoRA, and the motion denoiser runs in FP32.

Original vs Turbo

8.2× faster generation · 78% less peak process memory.

Same Mac: Apple M5 Pro, 48 GB RAM. Same 5-second motion and 100 diffusion steps. Fresh processes, warm disk cache, embedding cache disabled. Original: PyTorch CPU; Turbo: MLX GPU. Memory includes GPU allocations. Methods and precision checks.

Run

git clone https://github.com/octaviusp/kimodo-turbo.git
cd kimodo-turbo
./scripts/setup.zsh
.venv/bin/kimodo-turbo ui

Open http://127.0.0.1:8765, enter a prompt and click Generate motion. Requires macOS, Apple Silicon and the dependencies listed in the runtime README.

To download only the weights into an existing checkout:

.venv/bin/hf download Javascript/Llama-3-Kimodo-Turbo --local-dir out/weights

Licenses

NVIDIA Kimodo weights, Meta Llama 3 encoder and McGill LLM2Vec adapters retain their respective terms. See LICENSE and NOTICE.txt.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Quantized

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Javascript/Llama-3-Kimodo-Turbo

Base model

McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp

Finetuned

(3)

this model