rlx-models
A collection of 182 machine-learning model families spanning language, vision, speech and audio, ported as standalone Rust crates on top of the rlx multi-backend compiler and runtime, targeting CPU, Metal, MLX, CUDA and wgpu.
rlx-models provides concrete model graph builders and weight loaders for RLX, a multi-backend compiler and runtime, covering 182 model families across 16 categories in one crate per architecture. These include 36 language and multimodal LLMs such as Qwen3, Llama, Gemma and DeepSeek variants plus two diffusion language models, 12 vision-language and omni models, 10 vision encoders and detection or segmentation models, 5 biomedical models, image and video generation models, 15 automatic speech recognition models, 45 text-to-speech and speech-language models, 11 neural audio codecs, and 19 speech front-end, wake-word and DSP crates.
It is designed as a standalone repository meant to be cloned alongside the main rlx repository and built with Cargo as a workspace member, with an AGENTS.md file offering a quick reference for coding agents working in the codebase.
It is written in Rust and licensed under GPL-3.0.