State of Open Models Summer 2026 Report

TL;DR

Hugging Face’s Summer 2026 analysis reveals that Chinese labs now lead open‑source frontier model releases, while the majority of real‑world usage remains on sub‑1B models; Qwen has emerged as the de‑facto base model, and autonomous agents have become the Hub’s dominant traffic source.

1. Frontier model scale shifts to China

  • Chinese labs released the largest open models in almost every month of 2026, with parameter counts ranging from 754 B to 2.78 T.
  • In contrast, U.S. labs stayed below 130 B for five of seven months, except for NVIDIA’s Nemotron 3 Ultra (561 B) and Thinking Machines’ Inkling (952 B).
  • Labs such as Moonshot, MiniMax, Xiaomi, and Z.ai publish almost exclusively models > 70 B, targeting benchmark prestige rather than developer accessibility.
  • Tencent and Alibaba’s Qwen family spans the full size spectrum, providing both tiny and trillion‑parameter variants.
  • Hardware vendors (AMD, NVIDIA) dominate new model releases (> 200 repos each), using open models to showcase chip performance.
  • Most U.S. frontier releases above 100 B are adaptations of Chinese models rather than original creations.

2. Attention ≠ Adoption

  • The top‑25 repositories by downloads and by likes share only one common entry, demonstrating a clear split between hype and actual usage.
  • Downloads favor long‑standing, stable models (e.g., all‑MiniLM‑L6‑v2 with 1.55 B pulls) while likes spike for newly released frontier models.
  • Chinese frontier labs (MiniMax, Moonshot, DeepSeek, Z.ai) concentrate download volume on > 70 B models; U.S. labs (Google, Microsoft, IBM Granite) see almost no downloads for large models.
  • Qwen’s broad family achieved 2,045 M downloads in 2026, about 55× more than Moonshot’s frontier‑only portfolio (37 M).
  • Adoption is largely decided within a model’s first few months, after which usage plateaus.

3. Open‑weight licensing trends

  • Among 178 Chinese releases > 20 B parameters, 59 % use Apache 2.0 and 22 % use MIT; none impose non‑commercial restrictions.
  • U.S. large‑scale releases are less permissive: only 29 % are Apache/MIT, 41 % carry custom terms, and 30 % declare no license.
  • The permissive licensing suggests revenue is expected from API services, hardware sales, or platform positioning, not from the weights themselves.

4. Qwen becomes the community’s base model

  • Qwen derivatives total 151,448 repositories on the Hub, 2.6× Meta’s footprint and 4.7× Llama’s.
  • Daily creation rate: 180–210 new Qwen‑based repos per day during the first seven months of 2026.
  • Drivers of Qwen’s dominance:
    1. Consistent release cadence across many sizes.
    2. Broad coverage of use‑cases, allowing developers to stay within a single ecosystem.
    3. Apache 2.0 license, minimizing legal friction.
  • Most derivatives are community‑generated; Qwen itself contributed only 54 of the 28,531 GGUF conversions.

5. Small models remain the practical layer

  • Models < 1 B parameters account for **83 %** of all‑time downloads; > 100 B models capture 1 %.
  • In 2026, only 3 % of download volume went to models > 70 B.
  • The llama.cpp project, now fully supported by Hugging Face, enables local inference of trillion‑parameter models via GGUF formats.
  • GGUF downloads in July: 39.6 M for Qwen, 20.8 M for Gemma, 7.5 M for Llama.
  • Runtime‑related repositories (gguf, lerobot, Apple mlx) grew 464 %, 194 %, and 148 % respectively, outpacing core libraries like transformers (+16 %).
  • Labs could accelerate adoption by providing official GGUF conversions and clear quantization documentation, leveraging community converters such as Unsloth.

6. Agents are the new primary Hub users

  • The agent‑usage dataset (released July 2026) tracks token calls from autonomous agents to the Hub.
  • Claude Code dominated July traffic with 44.4 %, but its share fluctuated dramatically (67.8 % in April, 6.4 % in May), indicating a volatile, rapidly shifting market.
  • Approximately 25 % of agent traffic in July came from unidentified client identifiers, rising to 59.8 % in May, showing that new agents appear faster than registries can catalog them.
  • Recent infrastructure upgrades (agent‑trace datasets, agents.md endpoints, hf_fs tool, MCP integration into the Linux Foundation’s Agentic AI Foundation) reflect a strategic shift toward machine‑to‑machine interactions.
  • A documented autonomous intrusion in July used a quantized open model (GLM‑5.2) after closed‑model safety guards failed, underscoring emerging security considerations for agent‑driven workflows.

7. Looking forward

  • Geopolitical rebalancing continues: Chinese frontier models attract the most community attention, while U.S. contributions focus on hardware‑optimized conversions and broader‑scale families.
  • Small‑model dominance and the expanding local‑inference stack (llama.cpp, GGUF) will keep the ecosystem accessible, creating a feedback loop between developers, publishers, and downstream users.
  • As agents become the primary Hub traffic source, future reports will likely need new metrics and methodologies to capture machine‑centric activity.

Notes on method

  • Analysis covers Hub activity from January – July 2026.
  • Metrics (downloads, likes, derivatives, releases) reflect Hub‑specific signals and do not capture private deployments, API usage outside the Hub, or non‑Hub distribution channels.
  • Downloads measure usage within Hub pipelines; likes indicate community interest; derivatives signal downstream development effort.
  • Results should be interpreted as a partial view of the broader AI market.

Sources

Related