Open-Weight Model

Open weights vs. open source vs. proprietary

Three levels of openness are routinely conflated. A proprietary model (such as the GPT or Claude families) keeps its weights private - you can only reach it through the creator's hosted API, which means your prompts travel to their infrastructure. An open-weight model publishes the trained parameters under a licence, so anyone can run it themselves, but does not necessarily release the code and data needed to reproduce it. A truly open-source model goes further and publishes the training code and data as well. The Open Source Initiative's Open Source AI Definition draws exactly this line: weights alone are not "open source" without the data information and training code that let you study and rebuild the model.

In practice, almost every model marketed as "open source" today - including DeepSeek, Qwen, and Gemma - is really open-weight: the weights ship under a permissive licence (Gemma 4 under Apache 2.0, for example), while the training data stays private. The distinction rarely matters for using a model, but it matters enormously for who controls where it runs - and therefore where your data goes when it does.

Why open weights are the basis of sovereign inference

Because the weights are just a file of numbers, they carry no connection back to their creator. A weights file is a deterministic mathematical function: the same input produces the same output, with no mechanism inside it to send data anywhere. Whoever hosts the weights controls the entire data path. Call a proprietary API and the model creator's cloud sees every prompt; run an open-weight model on Infercom and the creator has published weights only, with no visibility into how they run - your data goes to our EU infrastructure instead. Think of published weights like a printed book: once the author releases it, they have no idea which library holds it or who is reading it.

This is what makes open weights the technical foundation for EU sovereign inference. The model's country of origin becomes irrelevant to where your data lives, because the creator is not hosting anything - a model trained anywhere can serve European users entirely under EU jurisdiction when it runs on European infrastructure. Requests to our EU-hosted models run in a Munich datacenter, on EU soil, under EU law, with no US parent company and no US CLOUD Act exposure. The regulatory framework recognises the category too: the EU AI Act sets lighter obligations for free and open-source general-purpose models than for closed ones.

Are open-weight models good enough?

The capability gap that once justified paying a premium for closed models has largely closed for everyday workloads. Independent trackers such as Epoch AI put the lag between the best open and best closed models at a matter of months, not generations - and note it is not widening. For most production tasks - chat, extraction, code, agents - a current open-weight model is competitive, and it can be served far faster on purpose-built hardware.

Running open weights also unlocks a dimension closed APIs cannot: control over where and how they run. On our dataflow infrastructure we serve open-weight models like MiniMax M2.7 and gpt-oss-120b at speeds well beyond typical GPU API providers, while keeping every request inside the EU. Origin becomes a question of behaviour, not data security - and the real question shifts from "can the model maker see my data" (architecturally, it cannot) to "do I trust my inference provider".

Related terms