This biannual analysis of the Hugging Face Hub ecosystem covers January to August 2026, drawing on downloads, likes, derivatives, and agent traffic to reveal how open-source AI is evolving.

A major finding is the geographical rebalancing at the frontier. Chinese labs such as Moonshot, MiniMax, Xiaomi, and Z.ai now release models from 754B to 2.78T parameters, often skipping smaller sizes entirely. American labs above 100B are rare: only Thinking Machines’ Inkling (952B), NVIDIA’s Nemotron 3 Ultra (561B), and a few others. Meanwhile, AMD and NVIDIA each published over 200 new model repositories, using open weights to drive hardware adoption. The center of gravity for frontier model creation has shifted from model labs to hardware and infrastructure companies.

Attention and adoption measure different things. Among the top 25 repositories by 2026 downloads and the top 25 by likes, only one appears in both lists. Models under 1B parameters take 83% of all-time downloads; everything above 100B takes 1%. The most downloaded model, all-MiniLM-L6-v2, was pulled 1.55 billion times with only 5,156 likes. Likes track excitement around frontier releases; downloads track stable infrastructure.

Licensing reveals the strategy behind open weights. Of 178 Chinese releases above 20B parameters, 59% carry Apache 2.0 and 22% carry MIT; none have non-commercial restrictions. American counterparts show only 29% under Apache/MIT. The return is not license revenue but API, hardware, and ecosystem positioning.

Qwen has become the community’s base model, with 151,448 derivatives on the Hub—2.6× Meta’s total footprint. Derivatives grow at 180–210 per day. Key factors: regular cadence, broad size coverage, and Apache 2.0 licensing. Downloads for Qwen reached 2,045M in 2026, 55× Moonshot’s 37M.

Small models remain the practical layer because most developers lack hardware for large ones. The gap is bridged by llama.cpp GGUF conversions: the July snapshot carries DeepSeek-V4-Flash at ~284B and Kimi-K3 at ~2.8T on consumer machines. GGUF downloads grew 464% over the period, far outpacing core libraries like transformers (16%).

A new dimension is agent traffic. For the first time, agent-originated requests can be tracked via agent/ tokens. Claude Code held 44.4% of July traffic, but its share dropped from 67.8% in April as Codex climbed to 20.8%. Nearly a quarter of agents remain unregistered, and new harnesses appear faster than registries can name them. In July, the Hub experienced its first documented autonomous-agent intrusion, analyzed using a quantized GLM-5.2 model after frontier closed models declined.

The report notes that Hub activity is one perspective—it does not capture API usage, private deployments, or other distribution channels. The ecosystem is reshaped every few months.