The AI ​​model makes open source, but right!

Thanks to the Deepsek, Meta, Mistral, Microsoft, Alibaba, Google and all others, publish the AI ​​model and their parameters under an open source license. However, open white models are not yet Open source aiBe brave, take an example Ai2 And publish all the data and the code with which you have made models! In my view, there are four good reasons.

Holgar Waymanman is a freelancer as a graduate informationist. For Heise online, he reports regular eclipse development on new release from the environment. He is an contributor to eclipse, llama.cpp and other open source projects.

IX Workshop: E-Billing Responsibility-InvioCa and ERP software adjustment

Four reasons for open source model

First, you are in the winning team. Surely you have some secret tricks in stock to train your model. But you have tried many, but not all options to find the best way. Type yourself together and bundle your innovative power. It took only four months and eight days Openaai o1 previewFirst argument model, not downloadable and with hidden argument, to independently available Deepsek R1This inspired the development of other logic models. In the worst situation, you only lose a lead for a short time.

Second, it saves resources: human resources, time and computing power. The latter is a real problem. Despite increasing efficiency during training and execution, the model usually requires more computing power than before as they are larger because more synthetic data is used during training and because current logic models often produce more tokens for logic. Your self -bullion causes unnecessary climate damage.

Third, it would be appropriate to publish training data only, as it is not your data. Code, for example, with which you train your model and it is necessary for logic, comes from open source projects. For many data, it is also not legally clear whether they can be used for training. And because you do not openly accept which data you use, yet there will be a long legal uncertainty.

Fourth, it will be helpful in use. With a look at the training data, you will be visually reduced by closing the eye and can try more specially: how do you format a table in an investigation? Markdown, HTML or latex? Are there a ruckus and indentation assistant, harmful or just waste of tokens? Google has recently published its GEMMA 3 model as an open weight. Want model Function calling is believed to be Permission; The one who is missing is a guide how it works. As someone says in the classic software development that the truth is in the code, it is in training data and in the training code for the AI ​​model.

For your new logic model QWQ-32B revealed AlibabaThe code generated during learning reinforcement by software tests was tested for purity. I can imagine well that an open source community can make a valuable contribution in covering more programming languages ​​or refining evaluation by quality matrix. We are especially large people, AI is still new and Python is not a favorite programming language, can support the next generation.

And another thing: Please publish not only your data and your code on the face and gitb, but also do it as an open source project in a manufacturer-independent location, for example in the Apache Foundation or Eclips Foundation. So it is more attractive for others to participate, and when a company is purchased, nothing disappears or is thinking differently. Open source is much higher for all of us than freeware. I hope you are there!

(May)