Global LLM Download Leaderboard: China models take 61.7% of top-300

The leaderboard's central design choice is its 30-day window. Cumulative download counts, the project argues, structurally favor older releases: a model published in 2022 can retain a large lifetime total even after users have moved on. By showing both the 30-day and cumulative figures on every row, the board makes that gap visible rather than asking readers to take the distinction on trust.

Its country attribution is equally deliberate. Models are grouped by lineage, meaning who trained them, rather than by the uploading organization, and third-party quantized files are credited back to the original model's lineage. The project says lineage is determined from model architecture rather than name, on the reasoning that a name can be changed but an architecture cannot.

The exclusions point in the same direction: automated test models are dropped because test harnesses, not people, drive their download counts, and embedding and reranking models are dropped because they do not generate text even though the platform often classifies them as text generation. The result is a board built to reflect human, generative use within a single month.

On the Korean side, VIDRAFT's models hold the top two spots and fourth place, accounting for 45.4% of downloads in the Korean top 30. POCKET-Darwin-180B-GGUF, a 180-billion-parameter reasoning model adapted to run on a laptop, was released on October 2, 2026 and drew 523,910 downloads in the past 30 days.