Just two years ago, most AI developers relied on closed, proprietary systems - especially for large language models (LLMs). Today, the landscape has flipped. Whether for text, image, audio, video, or fully multimodal tasks, a growing ecosystem of open models is evolving rapidly. On Hugging Face alone, over one million new repositories have appeared in the past 90 days. Making AI technologies openly available is key to building reliable tools and democratising their development: researchers and SMEs can adapt datasets and models without fearing that their tech stack will be deprecated or locked behind paywalls.
Open source AI (OSAI) models are getting better and better but also increasingly doing more with less compute. For example, models like Gemma 1B from Google enable complex tasks on edge devices with less than 1B parameters. These models are being actively downloaded, tested, and deployed across industries by the global community, with daily downloads hitting tens of millions for top repos.
The OSAI ecosystem continues to thrive, both in quality and quantity, with a significant shift in key players. NVIDIA has emerged as a surprising leader in this space, topping open source AI repository contributions in 2025. Their releases include the Nemotron family (agentic), BioNeMo (biopharma), Cosmos (physical), Gr00t (robotics), and Canary (speech recognition).
In this ecosystem, another big surprise is the spectacular rise of Chinese giants - led first and foremost by Alibaba Cloud and their beloved Qwen family of models, now powering everything from chat to multimodal reasoning and agent frameworks. Baidu, Tencent, MiniMax, Z.AI, ByteDance, Moonshot AI, and Zhipu AI are also closing in fast, matching or even surpassing Western contributors in repository activity and leaderboard performance. DeepSeek has over 100.000 followers on Hugging Face and is shipping updates of their V3 models.
Meanwhile, Europe is fading from the frame. Beyond Mistral AI’s efficient open models like Magistral, and Stability AI’s image generating models, few European firms are visible in this open source explosion, despite efforts for multilingual sovereignty. Expanding geographic diversity is one of the great challenges ahead for the AI community.
But clearly, this movement isn’t only about giants. Thousands of smaller teams and individual developers are uploading datasets, fine-tuned models, and task-specific variants every day, using accessible techniques like low-rank adaptation (LoRA) to customise base models with just hundreds of samples. Each contribution spreads understanding, builds literacy, and accelerates real-world AI problem-solving; from fine-tuned SLMs for on-device privacy to community datasets tackling niche domains like medical imaging.