Open Source LLM

Open source LLMs are language models whose weights — the trained parameters — are freely available. They can be downloaded, run locally, and adapted. That sets them apart from closed models, which are only accessible through the vendor’s API.

Difference from proprietary models

With proprietary models such as ChatGPT or Claude, the model stays with the vendor. All data flows through their infrastructure. Open models run wherever they are needed — including a company’s own data centre (self-hosted AI).

Well-known model families

Llama (Meta), Mistral (France), DeepSeek, and Qwen (Alibaba) rank among the most important open families. By 2026, the performance gap to closed frontier models has narrowed considerably.

Benefits and limitations

The benefits: full data control, no token fees when self-hosting, and adaptability through fine-tuning. The limitations: setup effort, dedicated hardware, and no vendor support — operations stay in-house.

Open weights does not mean open source

The licence decides. DeepSeek ships under MIT, Qwen mostly under Apache 2.0 — both genuine open-source licences without strings attached. Meta’s Llama, by contrast, comes with its own community licence including usage restrictions. Such models are correctly called “open weights”, not open source. Always check the licence before deployment.

As of: June 2026