Why develop locally
An agent makes many calls to the model for each task: it plans, calls a tool, checks the result and repeats. In development that is multiplied by hundreds of tests a day. With an open model on your own machine, every test costs the same: nothing.
The second reason is data. If the assistant will work with a client’s contracts, records or drawings, developing it locally avoids sending them to a third party and simplifies the confidentiality agreement.
What the computer has to load at once
In development there is never just one program open. The language model, a small model that turns documents into vectors, the vector database, the containers with the agent’s tools and the programming environment all run together. All of that needs memory and cores.
That is why unified memory fits well: the model uses the part it needs and the rest is left for everything else, without the fixed split of a dedicated graphics card. And a 16-core processor compiles and runs containers without the model suffering.
Which computer for each way of working
| Work | Computer | Why | Price |
|---|---|---|---|
| Prototypes and simple assistants | Compact AI 32 | Models up to about 19 GB and the whole environment around them | €2,065 |
| Agents with a large model | Compact AI 64 | Models up to about 42 GB, with memory left for the tools | €2,940 |
| Several models or several agents at once | Compact AI 128 | A large model to reason and small ones for quick tasks | €4,820 |
| Fine-tuning and training | Compact with RTX PRO or workstation | Training tools require an NVIDIA graphics card | Custom |
Tuning a model is not the same as using it
To use models, unified memory is enough. To tune them with your data, almost every tool works with CUDA, the NVIDIA technology. If fine-tuning is going to be part of the job, order the compact with an RTX PRO or consider a workstation with one or more graphics cards.
Several people on the same computer
A compact on the office network serves the whole development team: each person launches tests from their laptop and the model answers from the compact. Remote management over IP lets you restart it or enter the BIOS without touching it, even from home.
From the prototype to what the client receives
What works on your desk has to work the same at the client’s site. The safest way is to deliver the same machine you developed it on, with the same system and the same models. We build the series with your cloned image, your brand on the enclosure if you want it and a serial number per unit.
That is what our OEM programme covers: the same configuration in every unit, a 24-hour test and immediate replacement during the warranty.
Related hardware and guides
Frequently asked questions
Can I develop agents without an NVIDIA graphics card?
Yes. To run models and build agents, integrated graphics with unified memory is enough. An NVIDIA card is needed to tune or train models and for software that only works with CUDA.
Which operating system is best?
Linux is the norm in AI development: almost every tool is released for it first and the integrated graphics makes use of a little more memory. We deliver it installed free of charge. Windows 11 Pro on request.
Does the same computer work for the client demo?
Yes, and it is one of its advantages: it fits in a backpack, starts on its own and works without a connection. The demo runs on the same machine that is later installed.
Can you build several identical units for my clients?
Yes. We build the series with the same configuration, load your system image and deliver each unit with its serial number and test report. If you resell, the packaging can go without our brand.
What will you use it for?
Tell us which models you want to run, how many people will use it and where it will work. We will propose the compact, with or without a professional graphics card, and its price within 12 working hours.