Interconnects tracks USA vs China vs EU open-model downloads

The interesting question in open-source AI right now isn't whether a Chinese lab or a US one holds the current top spot on a benchmark, it's whose weights are actually being pulled down and built on. Nathan Lambert's Interconnects has now put a public number on that, with a daily-updated Open Models Dashboard showing Hugging Face downloads and derivatives of open-weight AI models across USA vs China vs EU.

The methodology is worth reading before quoting the numbers. The tracked models list on GitHub describes the scope as "Post-ChatGPT LLMs and VLMs (released after Nov 30, 2022)" with first-party weights on Hugging Face, new additions require over 100K total downloads, and guard/shield models such as Llama-Guard, ShieldGemma and Qwen3Guard are intentionally excluded so safety-classifier traffic doesn't inflate the base-model picture. The original seven organizations tracked, DeepSeek, Google, Meta, Microsoft, Mistral, mistral-community and Qwen, already cover 1,971 models through July 2025, and expanded coverage now sweeps in allenai, Arcee, NVIDIA, ByteDance-Seed, AIDC-AI, Apple, Cohere Labs, IBM Research and more than forty other orgs.

Why this matters even if you don't train models yourself: for the last couple of years there hasn't been a neutral, daily source of truth for whose open weights are actually winning. Vendor blogs cherry-pick and VC decks cherry-pick harder. A public dashboard with a documented cutoff date, download threshold and exclusion list is the kind of instrument that gets cited in policy rooms and enterprise pitches, and Interconnects has already been arguing, in The State of Open Models, that Chinese models overtook their counterparts built in the U.S. in the summer of 2025 on this exact download signal.

The honest caveat is that Hugging Face download counts are a proxy, not a measure of production usage, and the reporting doesn't tell you how a hybrid-geography lab is assigned to USA, China or EU, or how much of any number is CI mirrors versus real inference. Take the regional splits as reported, not settled.

Still, the direction is the interesting part. Whoever hosts the scoreboard tends to shape the narrative around it, and a methodology-first dashboard from a well-regarded post-training researcher is likely to become the reference sheet the next round of open-model policy fights is argued over.