Model Releases

  • LTX-2.5 Versatile Diffusion Model for Video and Audio — A diffusion model that handles bidirectional video, audio, and cross-modal generation in a single file. If you’re sick of juggling separate pipelines for image-to-video and text-to-audio, this cuts the glue code. 🤖
  • Qwen-Image 2.1 Uncensored Quantized Release — fp8/nvfp4 quantized version of Qwen-Image 2.1 that runs locally through ComfyUI or MLX. Lowers the memory bill for vision-LLM workflows without shipping your images to the cloud.
  • Qwen-Image 2.1 Turbo Uncensored GGUF Release — GGUF quantized version of Qwen-Image 2.1 Turbo that runs locally through ComfyUI or MLX. Enables efficient CPU/GPU acceleration for vision tasks without shipping data to the cloud.

Today’s Synthesis

Combining LTX-2.5’s unified cross-modal capabilities with the latency advantages of Qwen-Image variants creates a compelling local multimodal pipeline. Start by deploying LTX-2.5 for bidirectional video-audio generation where you want end-to-end coherence without managing separate pipelines—its single-file architecture eliminates glue code between vision and speech stages. Then layer Qwen-Image 2.1 Turbo Uncensored GGUF on top for rapid image-inference workloads within the same system; its GGUF format enables efficient CPU/GPU acceleration via tools like MLX or ComfyUI. This stack lets you run a complete visual-language-to-audio workflow entirely off-device: generate cross-modal media with LTX-2.5, then feed those outputs into fast-running vision components powered by Qwen-Image 2.1 Turbo. The key engineering win is avoiding round-trips to cloud APIs while maintaining state-of-the-art quality—ideal for edge deployments or privacy-sensitive applications requiring low-latency inference.