Knowledge · Local AI
Local AI use cases: what each one needs
Nine typical jobs for a language model in a small or medium-sized business, each with the hardware it needs, what that costs to buy or rent, how fast it runs and where the honest limit is. The figures are the same as in our calculator.
The nine use cases
From writing texts to AI agents
Graphics memory (VRAM) is the figure that decides. RAM and disk follow from it. Purchase means a complete computer; rental means a GPU server in continuous operation; by the hour it is much cheaper. All CHF figures are planning values, spot-checked on 17 October 2026, not quotes.
Writing texts and chat
8 GB VRAM · 16 GB RAM · 20 GB disk
Buy: CHF 900–1'400, one-off, suitable PC
Rent: CHF 260–450, per month for continuous operation; much cheaper by the hour
Speed: Smooth reading speed on a graphics card, good on a Mac, sluggish on the CPU alone.
Runs on an 8 GB graphics card (e.g. RTX 4060) or a Mac with 16 GB of memory or more, which many businesses already have. If need be, it even runs from main memory without a graphics card, though considerably slower.
Models: Qwen3.5 9B, Qwen3.5 4B
Speech to text (transcription)
8 GB VRAM · 16 GB RAM · 10 GB disk
Buy: CHF 900–1'400, one-off, suitable PC; the existing one is often enough
Rent: CHF 260–450, per month for continuous operation; much cheaper by the hour
Speed: Faster than real time with a graphics card: an hour-long meeting becomes text in a few minutes. On the CPU it takes about as long as the recording.
From 8 GB of graphics memory the large Whisper class runs briskly; smaller variants even run on the CPU alone. Macs are particularly well served here (whisper.cpp). Hardly any business needs to buy new hardware for this case.
Models: Whisper large-v3, Qwen3.5 4B
Generating images
8 GB VRAM · 32 GB RAM · 40 GB disk
Buy: CHF 1'200–2'500, one-off, suitable PC
Rent: CHF 260–450, per month for continuous operation; much cheaper by the hour
Speed: A few seconds per image on a good graphics card, minutes on a Mac or small cards.
From 8 GB of graphics memory (SD XL class); for smooth, high-quality results 16 to 24 GB is better (FLUX class). There is no way around a real graphics card here; Macs can do it, but considerably more slowly.
Models: Stable Diffusion XL, FLUX.1
Analysing PDFs and documents
12 GB VRAM · 32 GB RAM · 30 GB disk
Buy: CHF 1'500–2'500, one-off, suitable PC
Rent: CHF 260–450, per month for continuous operation; much cheaper by the hour
Speed: Answers at reading speed; very long documents take a moment to load first.
According to our hardware matrix the right card lies between 12 and 24 GB; 16 GB or a Mac with 32 GB or more is comfortable. Long documents mainly need context memory, which is why the larger card pays off here more than elsewhere.
Models: Gemma 4 12B, Qwen3.5 9B, GPT-OSS 20B
Connecting systems and internal search
12 GB VRAM · 32 GB RAM · 40 GB disk
Buy: CHF 1'500–2'500, one-off, PC or upgraded server
Rent: CHF 260–450, per month for continuous operation; much cheaper by the hour
Speed: Search hits practically instantly, written answers at reading speed.
A card of the 12 GB class or a Mac with 32 GB or more is enough to make company documents searchable and get answers with sources. Runs well on an existing server once a card is added.
Models: Qwen3.5 9B, nomic-embed-text
Excel & analyses
16 GB VRAM · 32 GB RAM · 50 GB disk
Buy: CHF 1'500–3'500, one-off, suitable PC
Rent: CHF 260–450, per month for continuous operation; much cheaper by the hour
Speed: At reading speed; complex analyses with large models take noticeably longer.
From 16 GB of graphics memory for the 14B class; for the considerably stronger 32B class you need 24 GB (e.g. RTX 3090/4090) or a Mac with 36 GB or more.
Models: Gemma 4 12B, GPT-OSS 20B, Qwen3.8 27B
Programming
16 GB VRAM · 32 GB RAM · 50 GB disk
Buy: CHF 1'500–3'500, one-off, suitable PC
Rent: CHF 260–450, per month for continuous operation; much cheaper by the hour
Speed: Suggestions within seconds, larger refactorings take time. Honestly: the cloud coding tools remain noticeably stronger.
From 16 GB of graphics memory for the 14B class as an entry point; local coding becomes really useful with the 30B coder class from 24 GB (e.g. RTX 3090/4090) or a Mac with 48 GB or more.
Models: Qwen3.5 9B, Qwen3-Coder 30B, Devstral Small 2, Qwen3.8 27B
AI agents
24 GB VRAM · 32 GB RAM · 60 GB disk
Buy: CHF 2'500–6'000, one-off, workstation; lower end only with a used card
Rent: CHF 300–450, per month for continuous operation, no maintenance of your own
Speed: A work step takes seconds to minutes; agents work in the background, so speed is rarely the bottleneck.
From 24 GB of graphics memory (e.g. RTX 3090/4090) or a Mac with 36 GB or more: agents need capable models and a large context, and the context eats memory. An agent runs continuously, so there is a lot to be said for a rented server rather than a computer running through the night.
Models: GPT-OSS 20B, Qwen3.8 27B
Generating videos
24 GB VRAM · 64 GB RAM · 80 GB disk
Buy: CHF 3'000+, one-off, workstation
Rent: CHF 450–1'400, per month for continuous operation, large GPU server
Speed: Minutes per clip, even on powerful hardware.
From around 24 GB of graphics memory for short clips (5 to 10 seconds is typical) in moderate quality, with minutes of waiting per clip. The honest recommendation: for most purposes the cloud is the better way.
Models: LTX-Video, HunyuanVideo
How to read this
Three rules behind the figures
Test before you buy
Whether a local model solves your task shows only with your own documents. Rent computing power by the hour for the test, then decide. Our accompanied two-week test does exactly that.
Cost is rarely the reason
At small volumes the cloud is cheaper. Most businesses that run models locally do so for data protection and independence, and that is a good reason even when local costs more. The cost calculator shows where the lines cross.
A Mac counts differently
On a Mac the model and the system share one memory, so the RAM figure is what counts there: 16 GB is the entry point, 32 GB is comfortable, 64 GB and more carries the larger models. A dedicated graphics card stays faster for images and video.
The ratings follow our internal Ollama hardware matrix (Q4 planning values, figures as of September 2026). More detail in German: the full German calculator page with model table, glossary and sources.
Your enquiry
Your message goes directly to Daniel Bidoli, the owner.
Address
Bidoli IT-Services
Feldeggweg 1
5033 Buchs AG