State of Open Models: Summer 2026 Observations

China Leads at the Frontier as US Hardware Firms Focus on Infrastructure

The open-source AI landscape has fundamentally shifted in 2026, with Chinese labs now consistently releasing the largest and most powerful models. While companies like Moonshot and Z.ai are pushing multi-trillion parameter models, U.S. contributions at the frontier are now focused more on optimization than creation. Hardware giants NVIDIA and AMD have become the most prolific American publishers, releasing tools and model variants that make massive models from abroad run efficiently on their chips, signaling a strategic realignment in the global AI ecosystem.

Downloads vs. Hype: A Split in the Ecosystem

An analysis of repository data reveals a stark disconnect between community attention and practical adoption. While the newest, largest models from Chinese labs garner significant attention, actual download volume is overwhelmingly dominated by smaller, older models that serve as dependable infrastructure. This highlights two distinct markets operating in parallel: a fast-moving frontier driven by benchmarks, and a stable infrastructure layer where most development work actually occurs.

• Downloads vs. Likes: The top 25 models by downloads and by likes in 2026 share only one model, indicating that popular models are not the ones most used in production pipelines.
• The 1% Rule: Models with over 100 billion parameters account for only 1% of all-time downloads, while models under 1 billion parameters command 83%.
• Dominant Base Model: Alibaba's Qwen has amassed over 151,000 community derivatives, 2.6 times Meta's total footprint, making it the de facto community standard.
• Permissive Licensing: 59% of large Chinese models use an Apache 2.0 license and none carry non-commercial restrictions, encouraging widespread adoption and modification.

The Rise of Qwen and the Power of Tooling

Alibaba’s Qwen has cemented its position as the community’s preferred base model, not just through its own releases but through a massive ecosystem of community-built derivatives. Its strategy of providing a full spectrum of model sizes under a permissive Apache 2.0 license has proven highly effective. This ecosystem is made practical by tooling like `llama.cpp`, which allows developers to run trillion-parameter models locally. This demonstrates that value is no longer just in the model weights, but in the surrounding tooling, community, and hardware optimization that makes these models usable.

The open-source AI competition is now a three-layer race: Chinese labs are winning the frontier model-size contest, U.S. hardware companies are winning the distribution and optimization battle, and community-adopted platforms like Qwen are winning the foundational ecosystem war. Direct monetization through licensing is dead; value now accumulates in the surrounding infrastructure and API services.