이미지: X — 모델·오픈소스
Summary
- In its "State of Open Models, Summer 2026" report released August 14, Hugging Face confirmed Qwen ranked first in local inference, with Gemma in second place
- The report noted that even as frontier models keep growing larger, smaller models are actually used more in practice
- The finding coincides with Qwen's early-August run — topping the Agentic Index and then releasing a 27B open-weight model
- 보고서명
- Hugging Face 'State of Open Models, Summer 2026'
- 발행일
- 2026-08-14
- 로컬 추론 순위
- 1위 Qwen, 2위 Gemma
- 보고서 핵심 관찰
- 프런티어 모델은 대형화, 실사용은 소형 모델이 주도
- 부가 관찰
- AI 에이전트가 허브 내 비중을 키우는 중
- Qwen3.8-Max 벤치마크(8/6 발표)
- Artificial Analysis Agentic Index 1위, Intelligence Index 5위(56점)
- Qwen3.8-27B 공개
- 8월 14일 아파치 2.0 라이선스로 허깅페이스 공개
The trend Hugging Face confirmed: Qwen leads in local inference
Hugging Face's "State of Open Models, Summer 2026" report, published on August 14, took stock of recent trends in the open-source AI ecosystem. The report pointed out that even as frontier models keep getting bigger, what's actually being used in the field is still small models. It also stated that in local inference — running models on-device rather than through cloud servers — Alibaba's Qwen ranked first, followed by Google's Gemma.
Local inference refers to running a model directly on a user's own PC or hardware rather than on cloud servers. Alibaba's Qwen team shared the news on its X account, briefly commenting that "small models make a big impact on the real world."
Why local inference matters as a metric
The Hugging Face Hub is where developers worldwide upload and download open-source model weights. Which models get downloaded and actually run locally the most is arguably a more honest measure of real-world usage than flashy demos. Running a model locally means data doesn't need to leave the user's system, and repeated usage costs shift from API fees to roughly the cost of electricity. The report also noted that AI agents — systems that perform tasks like search or file creation rather than simply generating answers — are taking up a growing share of activity on the Hub.
Timing overlaps with Qwen's recent moves
This result lines up closely with Qwen's recent announcements. On August 6, Alibaba's Qwen team said that Qwen3.8-Max ranked fifth (score of 56) on Artificial Analysis's Intelligence Index and first on the same organization's Agentic Index. Then on August 14, separate from its 2.4-trillion-parameter flagship model, Qwen released the 27-billion-parameter Qwen3.8-27B weights on Hugging Face under the Apache 2.0 license — a size built for easier local deployment. With Hugging Face's local inference ranking announcement landing the same day, the picture that emerges is Qwen building presence simultaneously at both the large-model and small-model ends of the spectrum.
| Metric | No. 1 | Note |
|---|---|---|
| Hugging Face Local Inference (8/14) | Qwen | Gemma ranked 2nd |
| Artificial Analysis Agentic Index (8/6) | Qwen3.8-Max | - |
| Artificial Analysis Intelligence Index (8/6) | Qwen3.8-Max | Ranked 5th, score 56 |
Editor's view
What makes this result interesting isn't the ranking itself, but how it was measured. Benchmark scores can be shaped by whatever definitions are chosen, but who actually gets downloaded and run locally on the Hugging Face Hub is closer to a vote cast by developers with their feet. Qwen leading on two fronts at once — the large-model race (Qwen3.8-Max topping the Agentic Index) and the small-model race (topping Hugging Face's local inference ranking) — suggests Alibaba is managing the "who's smarter" competition and the "who's used more" competition as separate tracks.
Anyone who has put open-source LLMs to work in practice will find this trend familiar. A few years ago, Llama was nearly the only viable option for a usable open model to run locally; over the past year or two, that spot has come to be shared between Qwen and Gemma. Running a 27B-class open model locally tends to lead to the same conclusion each time — it's not as capable as the latest large models, but it's more than adequate for repetitive, well-defined tasks like document summarization or code review, without incurring API costs.
For domestic companies and development teams, this is a practical signal. Not every task requires calling the top-performing model via API. For repetitive work, internal document processing, or tasks involving sensitive information, running a 27B-class open model on in-house servers may be the better choice on both cost and security grounds. That said, this applies mainly to well-defined tasks — complex reasoning or work requiring creative judgment is still better left to large models.
In the coming weeks, other open-model families are likely to challenge the download rankings on the Hugging Face Hub. The local inference race is only just heating up, and the next thing to watch is when a third name emerges to follow Qwen and Gemma.
