Run video, image and audio generation on hardware you own.
Every render leaves your studio and lands on someone else's GPU - metered, logged, and quietly used to train the next model. Sanctum runs the open models locally, so your work never leaves the room.
No API call to a public model provider
The cloud rents you your own creativity - by the render.
Image, video and assistant subscriptions, rebilled every month and metered on top. Locally, the box is the only cost.
Uploaded prompts and unreleased work feed provider training data and moderation logs by default.
Models get retired, throttled, or censored on the provider's schedule - mid-project.
A generation studio that lives on your desk, not in someone's data center.
Local video generation & upscaling
Text-to-video on Wan 2.2, HunyuanVideo 1.5 or LTX-2, plus frame interpolation and upscale passes - batched locally overnight.
Local image sets & style-consistent batches
FLUX.2, Stable Diffusion 3.5 and SDXL in ComfyUI - hold a locked style across hundreds of frames with your own LoRAs and ControlNet passes.
Local voice & audio (TTS, music, cleanup)
Narration with XTTS-v2 or F5-TTS, scoring with MusicGen, transcription with Whisper - voices and stems never touch a server.
Real hardware, real throughput.
Sanctum Mini
Apple Silicon · M4 Pro · 64 GB- Throughput
- 1 stream
- Seats
- 1
- Power
- 45–65 W
- Footprint
- 12.7×12.7 cm
- Noise
- Inaudible
Sanctum Max
Apple Silicon · M4 Max · 64 GB- Throughput
- 2-3 renders
- Seats
- 3-5
- Power
- 65–95 W
- Footprint
- 20×20 cm
- Noise
- Inaudible
The Mini and Max tiers run on an Apple Silicon MacBook Pro just as well as on a desktop - the same local stack, the same models, generating on the road with no connection required.
The open-weight video, image and audio models every workstation ships with - orchestrated in ComfyUI, running on the Apple Silicon sitting in your own studio. Open weights mean nobody can deprecate, throttle or censor them on their schedule.
Every render finishes inside the room it started in.
One box, the whole team. No metering, no upload, and no queue behind someone else's customers.
Read the case studyWhat creators ask before moving generation in-house.
How much VRAM do I need to generate video locally?
Video is the constraint, not images. LTX-2 generates comfortably on a 16GB card; the quality leaders - Wan 2.2 and HunyuanVideo 1.5 - want 24GB, though HunyuanVideo 1.5 will fit in roughly 14GB with offloading. 32GB covers every current image model at full quality, and above that the ceiling stops being VRAM and starts being how long you are willing to wait. On Apple Silicon the number to look at is unified memory rather than VRAM, which is why an M4 Max or M3 Ultra punches above its price here.
Is generating locally actually cheaper than a Midjourney or Runway subscription?
It depends on volume, and the honest answer is that a light user is better off on a subscription. The stack most working creators end up with - an image generator, a video generator, and two or three assistants - runs $70-120 a month and never stops. Local generation inverts the shape of the cost: the box is a one-time number and every render after it is free. The crossover is usually somewhere in the first year, and it arrives sooner the more you iterate, because iteration is exactly what metered tools charge you for.
Do AI providers train on my prompts and unreleased work?
Often, by default. Midjourney's terms allow images uploaded to Discord and used as image prompts to train future models. Across the wider market, prompts and uploads are commonly treated as training material unless you explicitly opt out, and the opt-out is frequently a paid enterprise tier. For unreleased client work or anything under NDA, that is a structural problem rather than a settings problem. Running the models locally removes it: nothing is uploaded, so there is nothing to retain.
Can this run on a MacBook, or do I need a desktop GPU?
A MacBook Pro with Apple Silicon works. The Mini and Max tiers run the same local stack and the same models on an M4 Max laptop as they do on a desktop workstation - the studio travels with you, and it keeps generating with the Wi-Fi off. The desktop tiers exist for throughput and concurrency, not because the software needs them: an Ultra or Cluster is what you buy when several people are generating at once or you are batching overnight.
Which open-weight models does it run?
For video: Wan 2.2, HunyuanVideo 1.5, LTX-2 and Mochi 1. For image: FLUX.2, Stable Diffusion 3.5, SDXL and Qwen-Image, with your own LoRAs and ControlNet passes. For audio: Whisper for transcription, XTTS-v2 and F5-TTS for voice, MusicGen and Stable Audio for scoring. Everything is orchestrated in ComfyUI, so the workflows you already have keep working. The models are open-weight and yours - nobody can deprecate, throttle or censor them on their own schedule.
Does it work with no internet connection?
Yes. Once the models are on the box there is no API call to make, so generation continues at full quality with the machine disconnected. Connectivity is only needed when you deliberately pull down a new model or update the stack.
See Sanctum generating on local hardware.
Twenty minutes, live. We run your kind of workload on the box and show you the numbers - no deck, no upsell.