Why this page matters for sovereignty
You can only run AI on your own hardware if you can download the model. Open weights are what make the Sovereign AI track possible at all.
What "Open Weights" Means
An open-weights model is one where the trained parameters — those billions of tuned sliders — are published for anyone to download and run. You can run it on your own hardware, inspect it, build on it, and swap it out whenever you like.
Contrast that with closed models (GPT, Claude, Gemini): you can send them prompts, but you can never download the model itself. The intelligence lives on the vendor's servers, under the vendor's rules, at the vendor's prices.
Open weights
- Download once, run forever — no per-token bill
- Runs where you choose: office, rack, private cloud
- Swap versions freely; no vendor can retire it
- You control updates, patching, and logs
Closed models
- Zero infrastructure to manage
- Usually the highest raw quality
- Pays per use, forever
- Vendor may deprecate or change it at any time
What "Open" Does Not Mean
Not open training data. You get the final sliders, not the textbook they were tuned from. How the model learned is not published.
Not automatically open-source. Many open-weights models ship under community licenses with strings attached — see below. "Open weights" describes availability; the license describes what you may do with them.
Not automatically safe. Open weights remove the vendor's guardrails-by-default. Running one responsibly is your job — which is exactly why policy ownership is part of sovereignty.
The Major Open Families
You'll see these names again and again. Family names change fast; the pattern is stable — a handful of major publishers, each releasing a ladder of sizes:
Llama (Meta) — the family that proved open weights could compete with the giants. Huge ecosystem of derived tools and fine-tunes.
Mistral (Mistral AI) — European, efficiency-focused; punchy small-and-mid models popular for self-hosting.
Qwen (Alibaba) — strong multilingual and coding line, a self-hosting favorite across many sizes.
Gemma (Google) — Google's open ladder, built from the same research line as its closed models; friendly sizes for laptops and workstations.
DeepSeek / Kimi — frontier-class open releases using mixture-of-experts designs to get big-model quality at lower running cost.
New releases land constantly. Judge any new model by the same three questions: How big? What license? Who trained it?
The License Fine Print
Permissive (Apache 2.0, MIT): use them for anything, including commercial products. The safest default for business use.
Community licenses (Llama, Gemma, Qwen variants): free for most uses, but with conditions — common ones are user-count ceilings (e.g., "free under 700M monthly users") and branding/attributions clauses. Fine for internal business use; read before embedding in a product.
Research-only or non-commercial: cannot be used for business purposes at all. Always check before deploying.
The three questions to ask about any model: How big? What license? Who trained it?