It’s been a hectic month in the upper tier of open-weight AI, with flagship models released by Qwen, MiniMax, Z.ai, Ant Ling, and StepFun. All eyes remain fixed on the imminent release of DeepSeek V4, as rumors continue to build momentum. Beyond the major frontier models, this problem is less severe across the long tail of specialized modalities and smaller model scales. Using our new Relative Adoption Metrics (RAM), we’re monitoring all these fresh releases by normalizing each model’s download counts against comparable models in the same size category. This has already proven to be an incredibly valuable tool for us, surfacing underrated models such as GPT-OSS — which is off the charts in terms of downloads and stands as the most popular American open-weights model since Llama 3.1. A RAM score above 1 indicates that the model is on pace to become one of the top 10 most-downloaded models of all time in its size category. We’re especially curious to see how early adoption of the smaller Qwen 3.5 dense models compares to that of Qwen 3 — balancing the rapidly growing Qwen brand against a more complex hybrid architecture that may strain the limits of certain open-source tools. Below is a summary of the RAM scores for several popular models released in late 2023, which clearly identifies Kimi K2 Thinking and certain OCR models as the standout performers. DeepSeek V3.2 and DeepSeek’s other recent large models have dramatically underperformed compared to the company’s earlier 2025 releases. (Time shown is days since release.) Qwen23-210B-A22026B by Qwen: The long-awaited update to the Qwen series has finally arrived.
Interconnects AI