Caravel AI
In development, not on saleWe are building an AI that runs on a server's processor and memory only. No GPU, so there is no GPU to rent or buy. It uses an existing open model that we run ourselves; we did not train it.
There is no form: you simply send us an email. There is no release date and no price yet.
CPU and RAM. No GPU.
No graphics card
The model loads into ordinary memory and the processor does all the work. The machine we tested on has no GPU at all.
No GPU to rent or buy
The hardware it needs is a server's processor and memory. Below, you can see what a GPU adds to a server's list price at one provider.
For work that can wait
It is slower than a live conversation. It suits drafts that a person reads and approves.
What a GPU adds to a server's price
Two dedicated servers from the same provider, Hetzner, with the same processor, memory and disks. The difference is the GPU. List prices read on 1 Nov 2026, excluding VAT.
| Server | EX44 | GEX45 |
|---|---|---|
| Processor | Intel® Core™ i5-13500 | Intel® Core™ i5-13500 |
| Memory | 64 GB DDR4 | 64 GB DDR4 |
| Disks | 2 × 512 GB NVMe | 2 × 512 GB NVMe |
| GPU | None | NVIDIA RTX PRO™ 4000 Blackwell SFF Edition, 24 GB |
| Price per month | €59.00 | €214.00 |
| Setup, once | €0.00 | €209.00 |
On these list prices, the server with a GPU costs €155.00 more per month, plus €209.00 setup.
What this comparison does not say
- These are server prices, not the total cost of running an AI. Operating work, monitoring and backups are not counted.
- This is not the price of Caravel AI, which has no price. We have not run Caravel AI on either of these servers; they are an example.
- A GPU can work faster, so a GPU server may complete more tasks for its price. We did not measure that.
- It does not mean cheaper than using AI through an API. By our own estimate of 1 Nov 2026, at low volume a European API costs less than a dedicated server without a GPU.
Sources: Hetzner, EX servers · Hetzner, GPU servers
What we measured
On 1 Nov 2026 we ran a small test that we wrote ourselves: 5 tasks with invented data, each run 3 times. The machine was a virtual server with 8 AMD EPYC vCPUs and no GPU, limited to 4 threads and busy with other work.
- Time per task: on average about 12 seconds with one model and settings, and about 45 seconds with another. These were the two highest-scoring setups.
- Memory: about 7 GB at peak in both.
- Long prompt with nothing cached: about 1 minute before the first word.
These numbers describe that test on that day. They are not a promise of speed or quality, and a test this small gives an indication, not a rate.
What it suits, and what it does not
It suits
- Drafts that a person reads and approves before anything is sent.
- Work where waiting seconds or a minute is acceptable.
It does not suit
- Live chat: tens of seconds per task is too slow for a conversation.
- Acting alone: in the same test, two of the three models that read a hidden instruction in a tool's text obeyed it and made a refund. The tools were simulated, so no money moved. A model's output cannot be the only safeguard.
Energy use: not measured
We have not measured the energy a task uses, with or without a GPU, so we make no claim about energy use or the environment. Lower power draw would not settle it, because a GPU can finish the same work sooner. Only measuring the same tasks on both can answer this.
Where it stands
- In development: we are working on making it faster and lighter, and we do not promise the result.
- Not on sale, and not something you can use today.
- No release date, no price and no final choice of model.
If this would be useful to you, tell us what you would use it for.