Why Chinese Open‑Source LLMs Overtook the US in 2025: Data‑Driven Insights

Background and Scope

The ATOM (American Truly Open Models) project released a comprehensive report analyzing the open‑source AI model ecosystem, focusing on download statistics, inference traffic, and derivative model development across the United States, China, and Europe.

Power Shift in the Open‑Source Landscape

Examining over 1,500 mainstream open‑source language models, the report identifies three clear phases of dominance:

Early stage (post‑ChatGPT): European models, especially Mistral AI’s 7B‑8x7B series, led the market.

2024: Meta’s Llama 3 series reclaimed U.S. leadership with high usability and varied sizes.

Summer 2025 onward: Chinese models (e.g., DeepSeek, Qwen) surged, overtaking U.S. models in downloads and inference share.

By March 2026, total open‑source model downloads reached 2.04 billion, a six‑fold increase year‑over‑year, with Chinese models showing the steepest growth (11.9× YoY, from 97 million to 1.15 billion downloads).

Download vs. Inference Metrics

While download counts measure static interest, inference token usage better reflects real‑world adoption. OpenRouter token‑share data shows Chinese models rising from 2.8 % to over 70 % of total inference traffic within 14 months, effectively swapping places with U.S. models.

Dominant Players

Qwen (Alibaba) emerged as the clear leader: 325 million downloads by September 2025, 942 million by March 2026, and a 69 % share of new derivative models by February 2026. Its success is driven by a dense release schedule and strong performance of small‑size models (10 B‑80 B parameters).

DeepSeek pursued a different strategy, focusing on ultra‑large expert models (>250 B parameters). It captured 47 % of historic downloads in the >250 B segment and commanded up to 75.6 % of inference traffic in June 2025.

Mistral peaked early with 57 % derivative‑model share in 2023 but fell behind, overtaken by DeepSeek in early 2026.

Meta’s Llama saw a rapid decline after Llama 4, dropping from a 37.4 % inference peak to near‑zero by August 2025.

Emerging Competitors

New entrants such as Nvidia’s Nemotron series, AI2’s OLMo, IBM’s Granite, and MiniMax added roughly 56 million downloads collectively, showing steady growth despite being far behind the giants.

OpenAI, traditionally closed‑source, entered the open‑source arena in September 2025 with the GPT‑OSS series, achieving 79 million downloads within months and quickly gaining significant inference share.

New Evaluation Metric: Relative Adoption Rate

To compare models of different scales, the report introduces a relative adoption rate that normalizes download counts against the median of the top ten models in the same parameter‑size bucket at fixed post‑release intervals (7 days, 14 days, up to one year). A score of 1.0 indicates average head‑model performance; higher scores signal exceptional “break‑out” popularity.

Key findings include:

Models in the 7‑9 B parameter range capture 33.8 % of total downloads, while sub‑100 B lightweight models account for 75 % of inference traffic.

Qwen 3.5 4B achieved a 5.29× score 14 days after release; DeepSeek’s OCR‑focused 3B model peaked at 9.42×.

OpenAI’s GPT‑OSS 120B reached 20.45× within a week and maintained 15.35× after six months.

Nvidia’s Nemotron Super 120B posted 18.03× and 19.93× in its first two weeks.

Conclusions

The open‑source AI ecosystem is undergoing rapid re‑centralization, with Chinese models now leading in both download volume and real‑world usage. Performance advantages, cost efficiency, and aggressive release cadences drive this shift, while traditional Western giants lose market share. The newly proposed relative adoption metric offers a more nuanced view of model popularity across parameter scales.

Data Sources

All download and inference statistics are derived from Hugging Face trends, supplemented by observations from China’s ModelScope platform.