With the emergence of generative artificial intelligence (AI), we hear more and more about open source models, opposing closed AIs like those of Openai, Anthropic or Google. The open approach is put forward more in Europe, as well as by the Chinese giants, here is what differentiates it from its rival.
Free models
Unlike proprietary models which often require costly licenses or subscriptions, open source AI can be downloaded, used and modified free by everyone, without license costs.
This means that researchers, developers and businesses of all sizes can use sophisticated AI capacities without financial obstacle (a priori). For example, a small business can use an open source language model to create a chatbot or a content generation tool without having to pay access to closed models like GPT-4O.
An AI that works locally
Being open source, the code of these models is accessible and can be executed on local machines without depending on an owner infrastructure or an internet connection. What drastically avoid the significant costs associated with the AI services hosted in the cloud.
This offers considerable advantages in terms of privacy and security protection, as organizations and individuals who exploit them can keep total control over their sensitive data, ensuring that they never leave their devices or internal networks. This approach is particularly valuable for sectors that process confidential information.
Local deployment also allows AI solutions to operate in areas where internet connectivity is limited, thus expanding the range of technology to isolated regions.
Big names have chosen open source
Deepseek, the Chinese startup that recently overthrew the AI sector with its R1 model, offers open technology. This is also the case of the French Mistral AI champion, and even Meta.
These models can be improved and corrected by the community
By making artificial intelligence models accessible to all, open source promotes collaboration between researchers, developers and businesses. Everyone can analyze algorithms, offer improvements and adapt models to specific needs, or even correct biases.
Rather than depending on a small number of private actors, innovation becomes collective and faster, each contribution benefits the whole community. The open source models also allow universities and startups, often limited in resources, to experiment with advanced technologies without prohibitive costs.
They revolutionize the daily life of small businesses
Thanks to their specificities, Open Source Code models transform the way companies approach artificial intelligence by offering profitable, customizable and transparent solutions. Unlike proprietary models, open source options allow companies to access, modify and freely deploy AI systems adapted to their specific needs.
For example, Meta Llama 3 has been largely adopted for tasks such as content generation and programming. Companies can refine it to create personalized chatbots or automate workflows without sacrificing data confidentiality by hosting it on their own infrastructure. Likewise, Bloom, developed by Hugging Face, supports 46 natural languages and 13 programming languages, which makes it ideal for global applications such as translation services or customer assistance.
Infrastructure challenges
Of course, this technology also has its limits. The infrastructure necessary to train the most advanced models is often out of reach of the general public or even many companies. They require, in fact, a colossal computing power, which makes them difficult to replicate without access to large -scale data centers.
Even if a pre-trained model is accessible, adapting it to specific use cases can require resources that only technological giants have.
In addition, fine-tuning, which consists in adjusting a model to improve its performance on specific tasks, requires advanced expertise and quality datasets. However, these are often protected or expensive to obtain. Likewise, exploiting these models effectively on less efficient equipment remains a major obstacle.
Safety risks
Another problem, Open Source AI poses major challenges related to cybersecurity; This is also the argument put forward by developers of closed models to highlight their technology.
With full access to algorithms, hackers are able to identify and use safety flaws, whether to biaise the results of a model or recover confidential information. For example, attacks make it possible to recreate a proprietary model by interacting with it, thus compromising its exclusivity and, potentially, the confidentiality of the data on which it has been trained.
Another risk concerns the integrity of data sets used to train these models. In the open source, where contributions are open to the community, an attacker could voluntarily introduce biases into training data, altering the behavior of the model. This technique can lead to discriminatory, offensive or even dangerous results.
Recently, Openai discovered that a Chinese entity operating Open Source led surveillance and influence in the world. One of his campaigns was to generate publications in foreign languages to criticize opponents of the government.
