Skip to content
NewsResearch

Alibaba's Qwen crosses 3 billion global downloads in six months, topping Meta and Google on Hugging Face

· by Pondero Newsdesk

The short version

Alibaba claims 3 billion+ downloads for the Qwen model family globally. A Hugging Face report published August 14 puts the on-platform gap in hard numbers: Qwen leads with ~2 billion downloads versus 418 million for Google and 227 million for Meta.

Alibaba's Qwen crosses 3 billion global downloads in six months, topping Meta and Google on Hugging Face

Alibaba says its Qwen open-model family accumulated more than 3 billion downloads worldwide in the six months through mid-August, per a company statement cited by Fortune. A Hugging Face State of Open Models: Summer 2026 report published August 14 put the gap in concrete terms: 2,045 million Qwen downloads on the Hub versus 418 million for Google and 227 million for Meta.

What

The Hugging Face report counted downloads within a year-to-date window, stripping any lifetime advantage for older models. By that measure, Qwen's 2026 downloads on the Hub reached 2,045 million across repositories with declared parameter counts (2,061 million including all repositories), against 418 million for Google and 227 million for Meta, per the report. Alibaba said the 3 billion total reflects downloads across all distribution platforms, not Hugging Face alone.

Qwen has released more than 460 open-source models, per Alibaba. On Hugging Face, 151,448 derivative models have been built on top of Qwen. That is 2.6 times Meta's total Hub footprint and 4.7 times the count for Llama repositories specifically, per the Hugging Face report.

Timed close to the report's release, Qwen also published Qwen3.8-27B on Hugging Face: a dense 27-billion-parameter vision-language model with a 262,144-token native context window, extensible to one million tokens via YaRN scaling, per the model card. The model handles image and video inputs, supports flexible reasoning control adjustable per request, and ships under an Apache 2.0 license on the Hub.

Why it matters

Download counts and likes track different things. The Hugging Face report drew the line explicitly: likes go to frontier models in the weeks after launch; downloads go to models already embedded in production pipelines running on a schedule. The top-downloaded models on the Hub are mostly from 2022. Qwen is the exception: it scores in both the attention and the adoption column, with 151,000-plus derivatives giving it the largest ecosystem base on the platform.

For operators choosing a base model for fine-tuning or agentic pipelines, that derivative count is practical signal. More derivatives mean more community fine-tuning recipes and pre-built quantized builds to choose from. The gap with Llama, currently 4.7 to 1 on derivative count, has compounding effects: each new Qwen fine-tune makes the ecosystem more useful for the next builder.

One licensing note from the Hugging Face report: the large-scale Qwen3.8 Max model (2.4 trillion parameters) is among the first Qwen releases to include non-commercial restrictions and revenue-share requirements. The 27B variant on the Hub carries Apache 2.0, but commercial builders scaling to the frontier tier should read the terms before deploying.

Context

Chinese AI labs have concentrated their Hugging Face volume in the large-model tier. The report found that effectively all of MiniMax's 2026 Hub downloads came from models above 70 billion parameters, and 88 percent of Moonshot's. By contrast, no major US lab records significant downloads above 70 billion parameters. Qwen's advantage is scope: the family spans every size range from the 2.4-trillion-parameter Max to the 27B dense model, and the report credited that breadth as the driver of its download lead over narrower frontier-only portfolios.

What to watch

Meta has not disclosed a Llama 4.2 release date. A Llama 4.2 launch or a Meta engineering post on ecosystem metrics would be the signal to watch: if Meta treats derivative count as a competitive measure, expect an accelerated open-source release cadence. Hugging Face's Summer 2026 report also flagged that its download data covers the Hub only. A cross-platform count that includes Modelscope and other distribution channels would either widen or contextualize Alibaba's 3 billion global figure.

Sources