Hugging Face State of Open Models: Summer 2026
A strong open-model ecosystem read about adoption, derivatives, small models, agent users, and why Qwen has become so important.
The model with the most likes on Hugging Face and the model with the most downloads are almost never the same. Attention and adoption measure different things, and conflating them is the most common mistake in open-model coverage.
Source
State of Open Models: Summer 2026 Observations
By: Hugging Face
Published: Aug 14 2026
Type: Report
This is one of the best ecosystem reads because it looks past individual model launches and asks what developers are actually building on. The data tells a different story from the benchmarks-and-hype cycle.
What the data says about where open models actually land
The headline numbers frame everything: 2.96 million public model repositories, 1 million datasets, 1.44 million Spaces. But the distribution is extreme — 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads.
Attention doesn’t equal adoption. Of the top 25 models by downloads and the top 25 by likes, exactly one appears in both lists. Likes go to frontier models in the weeks after launch. Downloads go to small, stable models wired into pipelines that run on a schedule. all-MiniLM-L6-v2 was pulled 1.55 billion times against 5,156 likes. Kimi-K3 got about 60 pulls per like. Likes tell you what the field is excited about; downloads tell you what it depends on.
The frontier is moving fast, but asymmetrically. In almost every month of 2026, the largest open model from a Chinese lab exceeded any model an American lab released. China’s monthly ceiling ranged from 754 billion to 2.78 trillion parameters; US models stayed under 130 billion in five of seven months (exceptions: NVIDIA’s Nemotron 3 Ultra at 561B and Thinking Machines’ Inkling at 952B). But the two biggest producers of new open model repos are AMD and NVIDIA — hardware vendors who’ve realised that open models optimised for their chips are the clearest proof the hardware works.
Qwen has become the community’s base model. Qwen-based derivatives total 151,448 repositories — 2.6× Meta’s entire footprint and 4.7× Llama specifically. Growth runs at 180-210 new derivatives per day. The position was built through consistency (regular release cadence, not occasional flagships), coverage (models from sub-1B to 2.4T parameters), and openness (Apache 2.0 licensing). Of the 28,531 GGUF conversions of Qwen models on the Hub, Qwen itself published only 54 — the community did the rest. Qwen gets 39.6 million GGUF downloads monthly, nearly twice Gemma’s 20.8 million and more than five times Llama’s 7.5 million, despite Llama having slightly more GGUF repositories.
Small models remain the practical layer. Models under 1 billion parameters take 83% of all-time downloads. Everything above 100 billion takes 1%. This held steady throughout 2026 — small models are the only ones that run on the hardware most developers actually have. But the llama.cpp project (whose team joined Hugging Face in February) moved the ceiling: the July snapshot carries GGUF builds of DeepSeek-V4-Flash at ~284B and Kimi-K3 at ~2.8 trillion. Local inference went from “8B on a laptop” to “trillion-parameter MoE across a few consumer machines.”
Agents are the new users. A new agent-usage dataset tracks coding agents calling the Hub. Claude Code led July at 44.4% (down from 67.8% in April), while Codex climbed from 10.4% to 20.8%. Nearly a quarter of agent traffic came from unregistered harnesses — new entrants are appearing faster than any registry can name them. Hugging Face responded by building for this reader: machine-readable papers, agent traces as datasets, agents.md endpoints on Gradio Spaces, and an MCP server exposing repos, docs, and papers in ~1,000 tokens.
Open weights shift where value accumulates. Of 178 Chinese releases above 20B parameters, 59% carry Apache 2.0 and 22% MIT. DeepSeek and Z.ai ship models up to 1.65 trillion parameters under plain MIT. The weights aren’t the product — the return comes from API/cloud business, hardware positioning, or ecosystem gravity itself.
My take
The finding I keep coming back to is the Qwen derivative count. It’s not that Qwen is the “best” model — it’s that Qwen covered every size band, shipped consistently, and used Apache 2.0. That combination created ecosystem gravity. Developers who start on Qwen for a small use case stay for a large one, and community-created derivatives reinforce the cycle. Meta built Llama’s brand through benchmark attention, but Qwen built adoption through shelf presence. The practical lesson for anyone choosing a base model: look at derivative counts and GGUF download volume, not benchmark headlines.
Where this is useful
- When choosing a base model for fine-tuning or deployment — derivative count and GGUF download volume are better adoption signals than benchmark rankings or likes.
- When deciding whether to release model weights — the data shows that permissive licensing (Apache 2.0/MIT) at all sizes, not just small models, drives ecosystem gravity. The value comes from the platform position, not licence revenue.
- When building tooling or infrastructure that agents will use — agents are already a measurable fraction of Hub traffic, and the interface they need (machine-readable endpoints, MCP tools, low-token discovery) is different from what human browsers need.
Apply this: Before picking a base model, check its derivative count on Hugging Face and monthly GGUF downloads. If the community isn’t building on it, you’ll be maintaining your fine-tune alone.
What I want to remember
Remember: For open models, ecosystem usage is the real value metric. The models many builders adopt as a base become infrastructure — and that position is built through consistent releases across sizes with permissive licensing, not through a single flagship launch.