When people say local AI they usually mean a chat assistant. In an office, the jobs that save the most time are often not chat at all: reading invoices, writing meeting minutes, searching documents, making a banner. Their memory needs are small, with one exception.
VRAM is the memory on the graphics card. For chat models we cover it in our guide to GPU memory for local LLMs. This post does the same for the other jobs. All figures are estimates from the AI selector's workload table, built from model cards and community guides. They are planning numbers, not measurements.
The one rule: size by the biggest single card
A language model can be split across cards, layer by layer, so two cards behave like one larger memory. Image, video and speech models generally do not work that way. One job runs on one card, and the model plus its working memory must fit there. A second card lets you run a second job in parallel. It does not make the first job bigger or faster.
So for these jobs, ask how big the largest single card is. Two 32 GB cards run two jobs at 32 GB, not one job at 64 GB.
Job by job
| Job | Example model | VRAM | Card class | Licence |
|---|---|---|---|---|
| Invoice and scan OCR | PaddleOCR-VL | about 8 GB | 16 GB | Apache 2.0 |
| Complex document OCR | DeepSeek-OCR 2 | about 16 GB | 16 to 24 GB | Apache 2.0 |
| Meeting minutes (Bahasa) | Whisper large-v3-turbo | about 4 GB | 16 GB | MIT |
| Document search (RAG) | bge-m3 | about 4 GB | 16 GB | MIT |
| Banners, product photos | Z-Image Turbo | about 12 GB | 16 GB | Apache 2.0 |
| Banners with text | Qwen-Image | about 20 GB | 24 to 32 GB | Apache 2.0 |
| Short video clips | Wan 2.2 A14B | about 32 GB | 32 GB | Apache 2.0 |
| Short video with sound | LTX-2.5 | about 32 GB | 32 GB | Check the licence |
RAG is retrieval-augmented generation: the assistant first searches your documents, then answers from what it finds. The search step uses an embedding model, which turns text into numbers that can be compared.
VRAM for the non-chat jobs, comfortable setting
Estimate. Lower fits more cards. Each job runs on one card.
- 16 GB card usable memory (15.4 GB)
- 32 GB card usable memory (31.4 GB)
Source: EMD Lab AI selector workload table, estimates. Not a measurement. Usable memory is the card's memory minus 0.6 GB for the driver.
What each group needs
OCR, speech and search are small. Whisper and bge-m3 fit in 4 GB. PaddleOCR-VL, a small model that reads invoices and tax documents, fits in 8 GB, and a quantised build (a compressed model file) needs far less. The selector notes that these jobs hardly affect the choice of GPU. Whisper supports Indonesian; one other speech model in the table, NVIDIA Parakeet, covers European languages only, so it is not the answer for Bahasa.
Images sit in the middle. Z-Image Turbo and SDXL are comfortable at 12 GB. Qwen-Image is about 20 GB and is the one the selector singles out for text rendering inside pictures, which matters for Indonesian marketing material.
Video is the heavy one. The smaller Wan 2.2 model lists about 24 GB comfortable, and the larger Wan 2.2 and LTX-2.5 list about 32 GB. On a 16 GB card these run only in heavily reduced form, with lower quality or with the model split between VRAM and system RAM, and they get much slower. If video is a real job, buy 32 GB on one card.
Camera QC (object detection on camera streams, for example counting items or spotting defects) is light. The selector lists the YOLO26 family at about 8 GB comfortable. Check its licence: the default one has conditions for closed-source commercial products.
Licences: check before you sell
Local does not mean free for every use. Some image and video models are free for commercial use only below a company-size limit, and some are non-commercial unless you buy a licence from the maker. In the selector's data, FLUX.2 [dev] and FLUX.1 [dev] need a commercial licence for business use, while the generated pictures may be used commercially. LTX-2.5 is free below a revenue limit. Qwen-Image, Z-Image Turbo, Wan 2.2 and PaddleOCR-VL are Apache 2.0. Read the licence on the model page for the exact version you use. Whisper and bge-m3 are MIT.
What to buy
- 16 GB covers OCR, Whisper, embeddings and a fast image model. It is the right size if chat and video are not the main job.
- 32 GB is the card for image and video work, and for a chat assistant at the same time. The standard Studio build has a 32 GB RTX PRO 4500.
- 96 GB lets you run several jobs next to a large language model without unloading anything: a chat model, Whisper, OCR and an image model, all resident. It also holds the biggest video settings.
- More than one card adds parallel jobs, such as one card for video and one for chat. It does not let a single video job use the combined memory.
Running several models at once needs the sum of their memory, plus working space for each. Add the figures in the table before you decide, and leave 10 to 20 percent free.
The short version
- OCR, Whisper and embeddings need 4 to 8 GB. A 16 GB card is plenty.
- Image models need 12 to 20 GB. Video wants about 32 GB on one card.
- Image and video jobs do not pool memory across cards. Size by the biggest single card.
- 96 GB is the card for running everything next to a large chat model.
- Check the licence of each model before commercial use.
All figures are selector estimates. See the model finder for chat models, or open the standard Studio build.