Since ChatGPT ignited the global AI wave at the end of 2022, big language models have undergone a huge transformation from closed source monopolies to open source prosperity. In the early days, OpenAI's GPT-4 and Google's Gemini represented the highest level of closed source routing, with astonishing capabilities but high API call costs, and model weights and training details were completely kept confidential. This closed source model has raised concerns in academia and industry about the "black box" nature of AI technology.
The turning point occurred in 2023 when Meta released the LLaMA series of models, pioneering the concept of "open source weighting". Subsequently, open-source models such as Mistral, Falcon, and Qwen emerged one after another, continuously narrowing the gap with closed source models in multiple benchmark tests. In 2024, DeepSeek's emergence became a milestone event in the open source ecosystem - its V2 and V3 models reached a level close to GPT-4 in reasoning, programming, and mathematical abilities, while training costs were only one tenth of the latter.
The success of DeepSeek is not accidental. The key technologies behind it include MoE hybrid expert architecture, Multi head Latent Attention, and GRPO reinforcement learning algorithm. These innovations not only reduce the training and inference costs of the model, but also provide a reproducible technical path for the entire open source community. The open source of the DeepSeeker R1 inference model has shocked the industry, as its inference ability has surpassed OpenAI's O1 model on multiple mathematical and programming benchmarks.
The rise of open-source models has had a profound impact on the AI ecosystem. Firstly, it significantly lowers the threshold for AI applications - small and medium-sized enterprises can deploy open-source models locally without having to pay expensive API fees or worry about data privacy breaches. Secondly, open source promotes technological democratization, allowing researchers worldwide to fine tune, improve, and innovate based on open source models, accelerating the iteration speed of AI technology.
Looking ahead to the future, open source and closed source models will coexist and promote each other in the long term. Closed source models still have advantages in efficiency and stability, while open source models are irreplaceable in terms of flexibility, transparency, and community innovation. The true winner of this open source revolution will be the entire AI industry ecosystem. [Reference source] This article is a comprehensive compilation of technical reports and open source community dynamics publicly released by major AI companies.