Open AI Model Ecosystem Now Runs on Far More Than a Handful of Chinese Labs

The open-weight AI model ecosystem is no longer dominated by a small clique of mostly Chinese labs. In a recent analysis tracking model releases, Interconnects found a broad and growing roster of organizations now contributing open models, spanning pure research houses, Big Tech giants, and product companies shipping specialized small models.

The source breaks makers into three buckets. "Pure" model builders — outfits like DeepSeek, Zhipu, Minimax on the Chinese side, plus Western firms such as Poolside, Arcee, and Zyphra, along with so-called sovereign AI players like Cohere, Mistral, and Trillion Labs — aim to push or approach the frontier. Big Tech participants play different games: Alibaba uses open releases to funnel users toward its closed offerings, and NVIDIA benefits because a thriving open ecosystem drives demand for its GPUs. Then there are product companies — JetBrains, Zed, Krea, Photoroom — that train small, specialized models to insulate themselves from dependency on closed API providers.

Recent releases illustrate the breadth. NVIDIA open-sourced Nemotron-3-Ultra-550B, a LatentMoE model it's licensing under the Linux Foundation's OpenMDW framework, replacing its prior custom license. Cohere released its flagship Command A+ — a 218B-parameter mixture-of-experts model — under Apache 2.0, a significant shift from earlier non-commercial terms. Poolside followed suit, releasing Laguna-M.1 under Apache 2.0 and pledging open weights as its default going forward. Zyphra, known for training on AMD hardware, shipped two new ZAYA1 mixture-of-experts models. And Zhipu's GLM-5.2 drew attention as a genuinely usable everyday model competitive with top closed offerings.

The analysis argues this diversity is a core structural strength of the open ecosystem, with teams reusing each other's training methods, architectures, and data recipes. Efforts to restrict or ban open model development would, in this view, simply concentrate power among a few closed providers — something the authors frame as both futile and dangerous for outside adoption of a transformative technology.