Open weights is not open source
Meta, Mistral, Alibaba, DeepSeek, Google and others release model files you can download, run, modify and, usually, build products on. This is called open weights, and it is a genuinely different thing from the closed models behind ChatGPT, Claude and Gemini.
It is also not what most programmers mean by open source. In almost every case the training data is not published, and often neither is the training code. You get the finished numbers, not the recipe. Some projects — AI2's OLMo is the clearest example — do publish data and code, and they are the exception worth knowing about.
Licences differ in ways that matter:
- Apache 2.0 or MIT. Do essentially what you like, including commercially. Most Qwen releases and several DeepSeek releases sit here, as do Mistral's smaller models.
- Community licences. Llama's, for instance, allows commercial use but adds an acceptable-use policy and conditions — historically including a clause that very large companies must ask for separate permission, and naming requirements for derivative models. Irrelevant to you personally. Very relevant if you are building a business on it.
- Research-only or non-commercial. Less common now, still around. Read before you build.
The short rule: downloadable does not mean unconditional. Open the licence file. It is two pages.
The families, briefly and honestly
Llama (Meta). The largest ecosystem — most tutorials, most fine-tuned variants, best supported by every tool.
Mistral (France). Efficient models with a strong reputation for doing a lot per gigabyte, and a European hosting story that matters for data-residency rules.
Qwen (Alibaba). Unusually strong across many languages, released in a wide range of sizes from tiny to very large, mostly under permissive licences. Often the best choice for non-English work you run yourself.
DeepSeek. Pushed the cost of strong reasoning models down sharply and released weights while doing it, which forced prices down across the industry.
Gemma (Google) and Phi (Microsoft). Small models built to run on modest hardware and punch above their size.
Where they are genuinely competitive
As of 2026, open models are close enough to frontier quality that the difference does not show for: summarizing, classification, extraction from documents, translation between well-resourced languages, rewriting, and coding help of everyday difficulty. If your task is one of those and you run it thousands of times, open models are usually the correct answer on cost alone.
Where they still lag: the hardest reasoning, long chains of tool use where the model has to plan across many steps, and broad world knowledge at small sizes. A 7-billion-parameter model simply knows less than a very large one, and it will invent things to fill the gap.
Why you would choose one
Cost at volume. Running an open model on rented hardware, or your own, can cost a fraction of frontier API prices for the same job.
Data that never leaves. Not a promise in a policy document. A structural fact about where the computation happens.
Permanence. Hosted models get retired. A model you rely on can be deprecated with a few months' notice, and your carefully-tuned prompts break. A file on your disk runs in 2030 exactly as it runs today.
Regulation. If your sector or country requires data to stay inside a border, this may be the only route open to you.
The catch
A downloaded model arrives with no safety layer, no moderation, no support, and no one to call. You are also now responsible for hosting, updates and the product layer around it. And "open" tells you about the licence, not about quality, bias or safety — those you still have to test.
The next lesson is the cheapest way to try one: your own laptop.
Before you move on