Models (and where your material goes)
Crust runs on your machine. Editing, compositing, layout, format variants, rendering and export happen locally and never leave your disk. When you run a cloud node, that one input is sent to the service hosting the model you picked, and the result comes back.
We route through seven services, Google under two separate agreements. This page says what each one’s terms allow them to do with your material, and where it is processed. Last verified: October 15, 2026.
| Service | What we run there | Trained on your material? | Rights to the output | Where it is processed |
|---|---|---|---|---|
| Anthropic | Claude text models, also used by Crustbot | No. Anthropic’s commercial terms exclude using API inputs and outputs to train its models. Held up to 30 days for safety purposes, then deleted. | You own the output. | United States |
| Google (Gemini API) | Gemini text models, Nano Banana image models, image description, Lyria music | No. We run on the paid tier, where Google’s terms exclude using your prompts, files or responses to improve their products. | Google claims no ownership of what you generate. | No fixed region |
| Google (Vertex AI) | Veo video | No. Vertex runs under Google Cloud’s terms, which include a training restriction on customer data, and their Cloud data processing addendum. | Google claims no ownership of what you generate. | United States |
| OpenAI | GPT text models, GPT Image 2 and GPT Image 2.5 | No. The API does not train on customer data by default, and every data-sharing setting on our organisation is switched off. Held up to 30 days for abuse monitoring, then deleted. | You own the output. OpenAI assigns its rights in it to you. | United States |
| SpaceXAI | Grok text models, Grok Imagine image and video | No. API access runs under their enterprise terms, which exclude using inputs or outputs for training. Deleted after 30 days. | You own the output. | No fixed region |
| ElevenLabs | Voice, speech-to-text, dubbing, sound effects, music | No. Training on our workspace data was switched off at account level on 21 September 2026. | Our plan includes a commercial licence and you keep ownership. Beta features and Music carry separate terms. | United States |
| fal.ai | Seedance, Kling, Happy Horse, Luma Ray, LTX, MiniMax H3, Topaz upscaling, Qwen edit, background removal, Recraft vectorize, Meshy, lipsync, reframe, segmentation, foley, Chatterbox, Sonilo, Demucs, SAM Audio | No, except Meshy 7.1 (see below). fal’s API terms exclude training on client content, including for the third-party models they host. Models they flag as not yet enterprise ready are carved out of that commitment, and we verify every model we run against that flag. | fal gives no warranty that output is original or free of third-party rights. Every model we run carries their commercial use licence. | United States |
| Alibaba Cloud Model Studio | Qwen Image and Wan video | No. Model Studio will not use your content to develop or improve its models unless you separately consent. | Model Studio claims no ownership of what you generate. | Stored in Singapore; processed in any region outside mainland China |
How we control which models run
We check every model we run through fal against fal’s enterprise-ready flag every day. If a model is not marked as enterprise ready, we are notified and switch it off, unless we have deliberately listed it as an exception. Today the only exception is Meshy 7.1.
Models hosted by fal but run by their owner
Many models available through fal are operated by the company that built them. For those, your material reaches that company as well as fal. Of the 52 models we run through fal, 37 work this way: all Seedance, all Kling, all Happy Horse, all Luma Ray, Sync Lipsync v1.9, v2 Pro and v3, both HeyGen models, Topaz image and video upscaling, Pixelcut and Bria background removal, Recraft vectorize, and Meshy 7.1 text, image, and multi-image.
Kling is Kuaishou, Seedance is ByteDance and Happy Horse is Alibaba, all based in China. The remaining 15 models run on fal’s own infrastructure, where fal’s terms are the whole protection: SeedVR, Qwen edit, smoretalk, LTX, Wan-VACE, SAM 3, Hunyuan Foley, Chatterbox, both Sonilo models, Demucs, MiniMax H3 Max, and the three SAM Audio separations.
Meshy 7.1 is the model fal currently flags as not yet enterprise ready, so fal’s no-training commitment does not cover those three endpoints.
Every model we run carries fal’s commercial use licence. None is marked research only.