State of Open Models: Summer 2026 Observations

In the hyper-accelerated timeline of artificial intelligence, a few months can feel like a decade. Following our biannual analysis of the ecosystem, we have synthesized the trends observed between January and August 2026. This report provides a deep dive into the data, revealing how the landscape of open-source AI is shifting beneath our feet.

The raw growth of the Hugging Face Hub remains staggering. Public model repositories have climbed from 2.43 million to 2.96 million, while datasets have surged from 711,000 to 1 million. Spaces, the interactive heart of the community, have expanded from 1 million to 1.44 million. Yet, beneath these headline figures lies a stark reality: the distribution remains heavily skewed. Approximately 85.6% of all models have fewer than 200 lifetime downloads, while a mere 1.5% of repositories account for 99.2% of all traffic. Every trend we discuss exists within this power-law distribution.

1. The Frontier is Moving Fast

The traditional progression of AI development—where labs start with small experiments and gradually scale up—has been disrupted. In 2026, several Chinese laboratories bypassed this incremental path entirely. Throughout the year, the most performant open models emerging from China consistently eclipsed the scale of their American counterparts.

China’s monthly ceiling for model parameters fluctuated between 754 billion and 2.78 trillion. In contrast, American labs largely remained under the 130 billion parameter mark, with notable exceptions like NVIDIA’s Nemotron 3 Ultra (561B) and Thinking Machines Lab’s Inkling.

Two Distinct Strategies

We are seeing a clear bifurcation in how labs approach the market:

• The Frontier-Only Portfolio: Labs like Moonshot, MiniMax, Xiaomi, and Z.ai rarely publish models below 70 billion parameters. They are betting on benchmark dominance and API demand.
• The Full-Spectrum Portfolio: Labs like Tencent and Alibaba (Qwen) release models across the entire size range, from sub-1B to massive multi-trillion parameter architectures.

The "Frontier-Only" strategy is now viable because size is no longer a unique differentiator. With companies like Xiaomi and Meituan clearing the trillion-parameter threshold, the barrier to entry for massive models has collapsed. Furthermore, the community’s rapid quantization layer ensures that even a massive model can be made runnable on consumer hardware within days of release.

Meanwhile, the U.S. presence in open source has shifted from software labs to hardware giants. AMD and NVIDIA are now the most prolific publishers, each releasing over 200 new model repositories this year. For these firms, open models serve as a strategic tool to demonstrate hardware capability. While Google and Meta were once the primary architects of the open-source movement, their focus has shifted—Meta, in particular, is increasingly prioritizing closed-source flagship models.

2. Attention ≠ Adoption

A critical insight from our 2026 data is the disconnect between "likes" and "downloads." When we analyzed the top 25 repositories by downloads versus the top 25 by likes, only a single repository appeared on both lists.

"A like says a release matters, and goes to frontier models in the weeks after they ship. A download says something is wired into a pipeline that runs on a schedule, and accrues to small, stable models over years."

This distinction is vital. Likes capture the industry's excitement, while downloads capture its infrastructure. For instance, the all-MiniLM-L6-v2 model recorded 1.55 billion downloads in just seven months, despite having only 5,156 likes. Conversely, frontier models from Chinese labs see high download volumes precisely because they are being integrated into large-scale production pipelines.

The data shows that Moonshot’s frontier-only portfolio recorded 37 million downloads, whereas Qwen’s broad-spectrum release strategy reached over 2 billion downloads. Coverage across use cases, not just raw power, is the primary driver of long-term adoption.

3. Open Weights Shift Where Value Accumulates

If frontier models were strictly a licensing business, we would expect the largest, most powerful models to carry the most restrictive terms. The data suggests the opposite.

Of the 178 Chinese releases above 20 billion parameters this year, 59% are licensed under Apache 2.0 and 22% under MIT. None of these massive models carry non-commercial restrictions. Even models ranging from 700 billion to 1.65 trillion parameters are being released under permissive licenses.

This indicates that these labs are not seeking direct licensing revenue. Instead, they are playing a long-term game:

• API and Cloud Integration: Driving traffic to their proprietary platforms.
• Hardware Positioning: Establishing their models as the standard for specific compute ecosystems.
• Ecosystem Dominance: Securing a "first-choice" position for developers, which creates significant valuation opportunities.

4. Qwen: The Community’s Base Model

By the metric of downstream impact, Qwen has cemented itself as the foundational layer of the open ecosystem. Qwen-based models now account for over 151,448 derivatives on the Hub—a footprint 2.6 times larger than Meta’s and 4.7 times larger than Llama’s.

Why Qwen Wins:

1. Consistency: A predictable release cadence that keeps the family relevant. 2. Coverage: Providing models for every use case, from edge devices to massive clusters. 3. Openness: Permissive Apache 2.0 licensing that removes friction for commercial builders.

The ecosystem is self-reinforcing: the broad model family attracts developers, who create derivatives, which in turn makes the ecosystem more attractive to future users. Remarkably, the vast majority of this growth is community-driven; Qwen published only 54 of the thousands of GGUF conversions currently in circulation.

5. Small Models Remain the Practical Layer

Despite the hype surrounding trillion-parameter models, the "practical layer" remains firmly rooted in small-scale models. Models under 1 billion parameters account for 83% of all-time downloads, while those above 100 billion parameters account for just 1%.

The bridge between these two worlds is llama.cpp. Since the ggml team joined Hugging Face, the project has become the most important tool for local inference. It has effectively moved the ceiling of what can be run on consumer hardware, allowing trillion-parameter Mixture-of-Experts (MoE) models to run on distributed consumer machines.

• Growth of Infrastructure: While model repositories grew by 21.5%, repositories declaring the gguf library rose by 464%.
• The Hardware Gap: Tools like Apple’s mlx and robot control stacks are growing three to seven times faster than standard transformer libraries.

Labs that fail to provide official, high-quality quantization artifacts are missing a massive opportunity to influence how their models are actually used in the wild.

6. Agents are the New User

The most significant shift in 2026 is the emergence of autonomous agents as the primary consumers of the Hub. In July, we tracked the agent/token traffic from coding agents calling our APIs.

The landscape is volatile. Claude Code led in July with 44.4% of traffic, but the market is characterized by rapid shifts—in May, a single client identifier accounted for nearly 60% of traffic. We are seeing new agentic harnesses arrive faster than we can register them.

This evolution has forced us to rethink our infrastructure:

• Machine-Readable Documentation: We now serve machine-readable Markdown for papers.
• Agent Traces: We have introduced agent traces as a first-class dataset type.
• MCP Integration: We have exposed repositories and sandboxes through the Model Context Protocol (MCP), allowing agents to interact with our ecosystem as if they were human developers.

In a striking development, we recently witnessed an autonomous agent attempt a sustained intrusion on our infrastructure. Interestingly, when we attempted to use closed-source frontier models to analyze the attack code, their safety guardrails prevented the work. We were only able to complete the analysis using a quantized open model, GLM-5.2, running on our own hardware.

Looking Forward

The geographical rebalancing of power in AI is accelerating. While the United States remains a hub for open-source development, the sheer velocity and scale of Chinese frontier models have captured the community's imagination and investment.

However, the AI race is a marathon, not a sprint. While tools like llama.cpp enable the local deployment of massive models, the true winners will be those who build a robust, positive feedback loop between developers, publishers, and end-users. As agents become the primary users of our infrastructure, the very definition of "adoption" is being rewritten. In the world of AI, the only constant is that the ecosystem will look entirely different in six months' time.

*

Note on Method: This analysis is based on activity observed on the Hugging Face Hub from January to July 2026. Metrics such as downloads and likes are indicators of ecosystem engagement and should not be interpreted as absolute measures of model quality or commercial market share.