Exe-Motion-Dynamic-V1-MLX

<div align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/exeterminal/Exe-Motion-Dynamic-V1-MLX/resolve/main/exe-chevron-dark.png"> <img src="https://huggingface.co/exeterminal/Exe-Motion-Dynamic-V1-MLX/resolve/main/exe-chevron-light.png" alt="Exe AI" width="72"> </picture> <h1>Exe AI</h1> <a href="https://exe-hq.net"><b>exe-hq.net</b></a> </div>

---

<div align="center"> <img src="exe-motion-dynamic-hero.png" alt="Exe Motion Dynamic V1 — local image-to-video for Apple Silicon" width="100%"> </div>

Exe Motion Dynamic V1

The video model of Exe Motion, built to run local on an Apple Silicon Mac. It is LTX-2.3 (22B distilled) with one change that matters on a 64 GB machine: the text encoder is quantized to 8-bit, so the part that used to be the memory spike of a run is now a quarter of its size. Website to the Main AI Agent Harness: https://exe-hq.net

Everything runs on-device through MLX — no cloud, no upload of your images or prompts.

What was quantized

<div align="center"> <img src="exe-motion-dynamic-quant.png" alt="Bar chart: the text encoder shrinks from 46 GB in F32 to 12.5 GB in 8-bit, four times smaller" width="80%"> </div>

The Gemma text encoder shipped in F32 and, at ~46 GB, was the biggest single thing a run had to hold. It is loaded, used to read the prompt, then freed before the transformer runs — so its size sets the peak. Quantizing it to 8-bit takes it to 12.5 GB, resident, without touching the transformer.

The transformer is left at full precision on purpose: the runtime has no quantized path for it, and a quantized copy would not load. The encoder is where the win is, and it is the honest one.

What it does

Image-to-video and text-to-video at up to 720p. Frame count drives the length (24 fps); resolution is free in steps of 64. A distilled fast path keeps steps low; there is no separate turbo weight to install.

Recommended settings

• Resolution: 1280×704 is the default; drop it for longer clips or to save memory.
• Frames: 8n+1 (81 ≈ 3.4 s, 121 ≈ 5 s). Longer needs more memory.
• Apple Silicon: long runs on a 64 GB machine still benefit from raising the Metal working-set limit before starting: sudo sysctl iogpu.wired_limit_mb=59392.

Files

A whole, assembled build — nothing to convert or borrow:

transformer/         22B distilled, bf16
text_encoder/        Gemma, 8-bit (this build's change)
text_projections/    connectors
vae/  audio_vae/  vocoder/  tokenizer/

License

Weights inherit the license of the base model, Lightricks/LTX-2.3. See the linked LICENSE for terms.