HuggingFace 近三个月趋势：大模型抢头条，小模型统治实际使用

数据来源：HuggingFace 夏季 2026 开放模型状态报告、agents-radar 日报归档、Tech AI Magazine 月度整理。

---

两件事同时在发生

过去三个月，HuggingFace 上有两条完全不同的叙事在同步推进。

上层叙事：前沿大模型持续刷新规模上限。GPT-OSS 120B（7月）、Gemma 4 多模态（8月）、GLM-5.2 MoE + Kimi-K3 长文档（9月）、DeepSeek-V4-Flash 低延迟推理——每个月都有”新高”。

下层叙事：实际被下载和使用的，绝大多数是你从没听说过名字的小模型。

HuggingFace 自己在夏季报告里给出了数字：小于 1B 参数的模型，占 Hub 全时段总下载量的 83%。大于 100B 的，只有 1%。

这不是说大模型没人用——而是说，当我们谈”AI 社区在用什么”，我们主要在谈小模型。

---

三个月热榜速览

7 月：前沿模型集体登场

7 月最大的事是开源/开权重大模型集体更新：

模型	参数	特点
openai/gpt-oss-120b	120B	首次开放权重
deepseek-ai/DeepSeek-V3.2	MoE	更新版本
moonshotai/Kimi-K2-Instruct	-	长上下文
zai-org/GLM-4.6	-	中文优势
Qwen/Qwen3-Coder-480B-A35B	MoE	代码专精
black-forest-labs/FLUX.1-Kontext-dev	-	图像编辑

同月，openai/whisper-large-v3-turbo（ASR）和 nomic-ai/nomic-embed-text-v2-moe（Embedding）也在榜。工具类模型和生成类模型并驾齐驱。

8 月：多模态与速度

Gemma 4 是 8 月的主角——Google 的全模态模型，覆盖文字、图像、视频和原生音频，从轻量边缘版到企业 MoE 版都有。DeepSeek-V4-Flash 主打低延迟推理，在开源社区里流行度持续上升。

9 月：MoE + 长文档

GLM-5.2 以混合专家架构重新进入热榜视野，Kimi-K3 以长文档处理能力（法律合同、学术论文、知识管理）获得关注。两者都对应具体业务场景，不只是 benchmark 数字。

10 月初（今日）：三个黑马

今天 agents-radar 日报里，前三名不是大模型：

Contrastive-LM/CLM-v0.1-8B（趋势分 485）
8B 参数的对比学习文本排序与验证模型。定位是 RAG pipeline 里的 reranker 和内容过滤层，不是生成模型。626 赞，2720 下载，属于”看起来没故事，但实际有用途”的那类模型。

TaichuAI/ZDTaichu5.0-9B（趋势分 475）
9B 视觉语言模型，专精空间推理——机器人、3D 场景理解、空间关系判断。2571 赞，12194 下载。中文团队出品，榜单上的稳定位。

nvidia/Nemotron-3-Diarization（趋势分 458）
100M 参数，专做说话人分离（最多 8 人）。M4 Mac 离线处理 10 分钟录音只需 3.1 秒。会议纪要、播客转写、客服录音标注的必备工具。26.4K 下载。

三个完全不同的细分任务，都是 10B 以下的专用模型。这是当日热榜给出的信号：开发者在为具体问题找具体工具，而不是找通用大模型。

---

小模型生态的真实规模

HuggingFace 夏季报告的另一组数字值得单独说：

GGUF 仓库增长 464%

相比之下，核心 transformers 库的仓库增长是 16%。GGUF 是专为本地推理设计的量化格式——这个增速意味着，“让模型在本地能跑”的需求，正在以压倒性的速度增长，远超新模型的发布速度本身。

unsloth/Qwen3.8-27B-GGUF 一个仓库就有超过 1100 万次下载。这不是在下载”开源大模型”，这是在下载”在我的机器上能实际运行的量化版本”。

Qwen 的社区主导

Qwen 系列在 Hub 上衍生出超过 15.1 万个模型，是 Meta（Llama 系列）衍生数量的 2.6 倍。Qwen GGUF 格式每月下载量 3960 万次，是 Llama GGUF 的 5 倍以上。

这意味着 Qwen 已经不只是一个模型系列，而是当前开源社区微调和量化的默认底座。当开发者想做一个垂直领域的小模型，第一个想到的基座大概率是 Qwen。

---

两套逻辑，同时成立

理解当前开源模型生态，需要接受两套逻辑同时成立的现实：

发布侧的逻辑：参数量和 benchmark 分数是竞争维度。大模型公司需要用每次新发布证明自己仍在前沿，GPT-OSS、Gemma 4、Kimi-K3 都是这个逻辑下的产物。这是新闻，也是必要的技术进步。

使用侧的逻辑：能不能在我的机器上跑，是不是能解决我的具体问题，模型有多小、推理有多快。83% 的下载量在 1B 以下，464% 的 GGUF 仓库增长，都在说同一件事：大多数开发者在意的是可用性，不是能力上限。

今日热榜的 CLM-v0.1-8B 是一个典型——没有参数大战，没有 benchmark 刷新，只是一个在 RAG pipeline 里做 reranking 的 8B 专用模型，却在当天排在趋势榜前三。这种模型不会上技术媒体的头版，但它会被几千个实际在做产品的开发者安静地下载和部署。

---

要关注的几个细分方向

根据三个月的热榜，以下几类模型反复出现：

RAG 工具链：reranker、embedding 模型持续上榜，说明 RAG 应用规模在扩大，工具链需求被具体化了。

音视频处理：Nemotron-3 说话人分离、Whisper turbo、腾讯 AuK 语音——实时/离线音频处理需求形成了稳定的细分赛道。

空间和多模态推理：ZDTaichu5.0-9B 的上榜是一个信号，空间理解（不只是图文问答）开始成为独立的能力评估维度，背后可能是机器人/工业自动化的需求在驱动。

Qwen 衍生微调：量化和微调的 Qwen 变体，包括各种语言、各种垂直领域，是社区活跃度的主要贡献来源。

---

关键数字汇总

指标	数据
<1B 模型占总下载量	83%
>100B 模型占总下载量	1%
GGUF 仓库增长（年化）	+464%
transformers 仓库增长	+16%
Qwen 衍生模型数量	151,448
Qwen vs Meta 衍生比	2.6x
Qwen GGUF 月下载	3960万次
Qwen vs Llama GGUF 比	5x
HuggingFace 数据集总量	100万+
85.6% 模型的下载量	<200次

---

数据来源：HuggingFace 夏季 2026 开放模型状态报告（huggingface.co/blog/state-of-open-models-summer-2026）、agents-radar 日报、Tech AI Magazine 月度整理。以上仅供技术趋势参考，不构成模型选型建议。

---

HuggingFace Q3 2026 Trending: Big Models Make Headlines, Small Models Run the World

Data sources: HuggingFace Summer 2026 State of Open Models report, agents-radar daily digests, Tech AI Magazine monthly rankings.

---

Two Stories Running in Parallel

The past three months on HuggingFace have been driven by two completely separate narratives.

The headline narrative: Frontier models keep pushing scale ceilings. GPT-OSS 120B (July), Gemma 4 multimodal (August), GLM-5.2 MoE + Kimi-K3 long-context (September), DeepSeek-V4-Flash — each month had a new “record.”

The usage narrative: What’s actually being downloaded and run is almost entirely models you’ve never heard of — and most of them are small.

HuggingFace’s Summer 2026 State of Open Models report gives the number directly: models under 1B parameters account for 83% of all-time downloads on the Hub. Models over 100B account for 1%.

This isn’t a story about big models being irrelevant. It’s a story about what “the AI community” actually uses every day.

---

Q3 Trending Summary

July: Frontier model wave

GPT-OSS 120B (first open-weights from OpenAI), DeepSeek-V3.2, Kimi-K2-Instruct, GLM-4.6, Qwen3-Coder-480B-A35B (code-specialized MoE), FLUX.1-Kontext-dev (image editing), whisper-large-v3-turbo, nomic-embed-text-v2-moe. Tool-class models trending alongside generative ones.

August: Multimodal and speed

Gemma 4 dominated — Google’s full-modality model handling text, image, video, and native audio, from lightweight edge variants to enterprise MoE. DeepSeek-V4-Flash gained traction for low-latency inference.

September: MoE and long documents

GLM-5.2 re-entered attention via MoE architecture; Kimi-K3 for long-context enterprise use (legal contracts, academic literature, knowledge management). Both tied to concrete business scenarios, not just benchmarks.

Early October (today): Three surprise entries

Today’s top three on agents-radar aren’t large models:

Contrastive-LM/CLM-v0.1-8B (trend score 485): 8B contrastive learning model for text ranking and verification. Built for RAG pipelines as a reranker and content filter — not a generative model. 626 likes, 2720 downloads. The “no narrative but actually useful” category.

TaichuAI/ZDTaichu5.0-9B (trend score 475): 9B VLM specialized for spatial reasoning — robotics, 3D scene understanding, spatial relationship judgments. 2571 likes, 12194 downloads. Chinese team, consistent trending position.

nvidia/Nemotron-3-Diarization (trend score 458): 100M parameters for speaker diarization (up to 8 speakers). M4 Mac processes 10 minutes of audio offline in 3.1 seconds. 26.4K downloads.

Three completely different specialized tasks, all under 10B parameters. The signal: developers are looking for specific tools for specific problems, not general-purpose large models.

---

The Real Scale of the Small Model Ecosystem

GGUF repository growth: +464%

Core transformers library repos grew 16% over the same period. GGUF is the quantization format designed for local inference. This growth rate means “making models actually run locally” is growing at an overwhelming pace — far outpacing even the release of new models.

A single repo — unsloth/Qwen3.8-27B-GGUF — has over 11 million downloads. Developers aren’t downloading “a large open-source model.” They’re downloading “the quantized version that will actually run on my hardware.”

Qwen’s community dominance

Qwen has over 151,000 derivative models on the Hub — 2.6x Meta’s (Llama) total footprint. Qwen GGUF downloads reach 39.6 million per month — 5x Llama GGUF.

Qwen has become the default fine-tuning and quantization base for the open-source community. When a developer wants to build a vertical-domain small model, their first instinct is now Qwen.

---

Two Logics, Both True

Understanding the current open model ecosystem requires accepting that two things are simultaneously true:

The release side’s logic: Parameter count and benchmark scores are the competitive dimension. Big model labs need each release to demonstrate continued frontier relevance. GPT-OSS, Gemma 4, Kimi-K3 all live in this logic. This is news, and it represents genuine technical progress.

The usage side’s logic: Can it run on my machine? Does it solve my specific problem? How small is it, how fast does it run? 83% of downloads under 1B, 464% GGUF growth — both point to the same conclusion: most developers care about deployability, not capability ceiling.

CLM-v0.1-8B today is a case in point — no parameter competition, no benchmark record, just an 8B specialized reranker for RAG pipelines, quietly ranking top-3 on the trending digest. This kind of model won’t make tech media front pages. But it will be silently downloaded and deployed by thousands of developers actually building products.

---

Segments to Watch

Three months of trending data points to consistent signals in specific categories:

RAG toolchain: Rerankers and embedding models keep appearing. RAG applications are scaling up, and toolchain demand is becoming concrete and specialized.

Audio/video processing: Nemotron-3 diarization, Whisper turbo, TencentAuK speech — real-time and offline audio processing has become a stable niche with consistent demand.

Spatial and multimodal reasoning: ZDTaichu5.0-9B’s appearance is a signal that spatial understanding (not just image-text Q&A) is becoming an independently evaluated capability — likely driven by robotics and industrial automation demand.

Qwen derivative fine-tunes: Quantized and fine-tuned Qwen variants across languages and vertical domains are the primary contributor to community activity volume.

---

Key Numbers

Metric	Value
<1B models share of total downloads	83%
>100B models share	1%
GGUF repository growth	+464%
Transformers library growth	+16%
Qwen derivative models	151,448
Qwen vs Meta footprint	2.6x
Qwen GGUF monthly downloads	39.6M
Qwen vs Llama GGUF	5x
HuggingFace total datasets	1M+
Models with <200 downloads	85.6%

---

Data from the HuggingFace Summer 2026 State of Open Models report, agents-radar daily digests, and Tech AI Magazine monthly rankings. For technical trend reference only.