Chinese Frontier Models, Qwen and AI Agents Reshape Hugging Face
Hugging Face finds Chinese frontier models and Qwen reshaping open AI, as local inference and coding agents surge while downloads remain concentrated.
Summary
On August 14, 2026, Hugging Face reported that, during 2026’s first seven months, public model repositories rose from 2.43 million to 2.96 million, datasets from 711,000 to 1 million and Spaces from 1 million to 1.44 million. Yet 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories capture 99.2%. Chinese monthly frontier ceilings ranged from 754 billion to 2.78 trillion parameters; original U.S. models above 100 billion were Thinking Machines Lab’s Inkling at 952 billion, NVIDIA’s Nemotron 3 Ultra at 561 billion and Nemotron 3 Super at 124 billion, and Arcee AI’s Trinity-Large at 399 billion.
Of 178 Chinese releases above 20 billion parameters, 59% use Apache 2.0, 22% MIT and none are noncommercial; comparable U.S. shares were 29% Apache or MIT, 41% custom and 30% undeclared. Only one repository overlapped the download and like top 25; no 2026 model entered the download ranking, while 13 came from 2022 and all-MiniLM-L6-v2 logged 1.55 billion pulls against 5,156 likes. Moonshot drew 37 million downloads versus Qwen’s 2.045 billion declared, 2.061 billion overall. Qwen has 151,448 derivatives, 2.6 times Meta and 4.7 times Llama, growing 180 to 210 daily; Google has 82,506.
Models under 1 billion parameters take 83% of all-time downloads, models above 100 billion take 1%, and those above 70 billion receive 3% in 2026. llama.cpp now carries roughly 284 billion DeepSeek-V4-Flash and 2.8 trillion Kimi-K3 builds; monthly GGUF downloads are Qwen 39.6 million, Gemma 20.8 million and Llama 7.5 million. July agent traffic was 44.4% Claude Code, after 67.8% in April and 6.4% in May, while Codex rose from 10.4% to 20.8%; unnamed harnesses supplied nearly one quarter in July and 59.8% in May. After an autonomous intrusion, closed models refused analysis, but GLM-5.2 completed it. Hub metrics exclude APIs, private deployments and external distribution.
Positives
- AMD and NVIDIA each released more than 200 model repositories, while LiquidAI ranked third with roughly 100.
- Qwen’s 151,448 derivatives include 28,531 GGUF conversions, only 54 published by Qwen, demonstrating extensive community participation.
- The ggml team joined Hugging Face in February while llama.cpp remained fully open source and community governed.
- GGUF repositories grew 464%, LeRobot 194% and Apple’s MLX 148%, outpacing model repository growth of 21.5%.
- March brought machine-readable papers, April added agent traces and agents.md, and July introduced hf_fs and secure attachable sandboxes.
- DeepSeek and Z.ai released models from 700 billion to 1.65 trillion parameters under the permissive MIT license.
Risks & concerns
- 1.5% of repositories account for 99.2% of downloads, while 85.6% of models remain below 200 lifetime downloads.
- Most U.S. releases above 100 billion parameters build on Chinese models, and AMD produced conversions but no original model at that scale.
- Moonshot, MiniMax, Xiaomi and Z.ai publish almost nothing below 70 billion parameters, putting their models beyond most developers’ hardware.
- No model published in 2026 reached the download top 25, showing attention for new frontier releases has not yet translated into established usage.
- Claude Code’s agent share swung from 67.8% in April to 6.4% in May, showing that defaults or releases can abruptly shift traffic.
- An autonomous agent conducted a sustained intrusion against Hugging Face, while frontier closed models declined to analyze the captured attack code.