Field sheet 01 · Chapter 11 · 2 October 2026

Host your own AI

on hardware you control

Run the model yourself and nothing leaves the building. No data policy to trust, no terms that can change, no account shut off mid-session. Until recently this meant a much weaker model. Not any more.

download the PDF read the full chapter

The Artificial Analysis Intelligence Index by open weights and proprietary, the 25 models it shows by default, in rank order with their names and scores. — Fig. 01The 25 models Artificial Analysis shows by default on its Intelligence Index, 2 October 2026, ranked. Twelve of the 25 publish their weights. The best open one, MiMo-V2.6-Pro from Xiaomi, is tenth with 46, twelve points behind Claude Opus 5.5. The index is now version 4.3.2, with ten tests, and its scores run lower than in August, so the two sets cannot be compared. Source: artificialanalysis.ai/#intelligence.

01Why run it yourself

Your data stays with you. National positions, instructions from your capital, and anything said in confidence never reach an outside company.
Nobody can raise the price or switch the model off in the middle of a session.
You can see every step. Online services hide the model's reasoning. On your own server you can log it and show a colleague.
Your institution builds the skill instead of renting it every year.

02What it costs

One NVIDIA DGX Spark, 128GB. A mid-sized model for sixty people at once.$4,699

Two units linked directly, 256GB. A large model like DeepSeek, faster than you can read, holding thousands of pages at once.$9,400

Three or four units, 384 to 512GB. The largest open models, in full.$14k to $19k

Compared with a subscription for every delegate, the hardware pays for itself within a cycle or two, and keeps working after the funding ends.

03What to run

Model	Licence	Index
MiMo-V2.6-Pro	Open	46
GLM-5.3	Open, restricted	45
Kimi K3	Open, restricted	44
GLM-5.3-Flash	Open, MIT	42
Qwen3.8 2.4T	Open, restricted	40
DeepSeek V4.1 Flash	Open, MIT	39
Qwen3.8 27B	Open	34

Read the licence before you choose. Several of the strongest open models limit commercial use. That is usually fine for a delegation, but not for a company building a product on one.

Run it with vLLM or SGLang. Both work like OpenAI's service, so any tool that works with ChatGPT works with your server once you change the address and key. On a laptop, ollama run qwen3.8:27b

04Rented or sovereign

Size or local control. A shared system is cheap and strong. It also decides the categories, makes you depend on it, and can make leaving impossible.
Power and risk. Whoever holds the power should carry some of the risk. When you rent, the provider writes the terms and you live with the result.
Open does not mean trusted. Closed systems can be checked under contract. You need a lasting right to check, and holding the weights is the cheapest way to get it.
Skills build on each other. New open models come out every few months. A team that runs its own tries each one within a week and keeps the know-how. A team on a contract waits.
The right to stay out. Some knowledge belongs in no model. Sovereignty includes refusing to become data.

From the AI4PG workshop on Decision Sovereignty in Nature, ETH Zurich, July 2026.

05Before you trust it

The index averages ten tests, and open models do not fall behind evenly. The gap is biggest on the hardest reasoning tasks and on hallucination, where closed models are still ahead.

For a negotiator, hallucination matters most. Point the model at your own documents, ask for sources, and open them.

06If you cannot yet

Use an online service that promises in writing to store nothing you send. Treat it as a temporary fix. Online models for public material and practice. Your own hardware for any position not yet public.