Hugging Face Top 10: Google EmbeddingGemma 2 jumps straight to No. 1

Every Monday I pull the public Hugging Face trending ranking, sorted by trendingScore, and write up the Top 10. This is week 2026-W42, the second snapshot in the djcroman series. Last week set the baseline, so this is the first week where I can show who climbed, who slipped and who dropped out of the list completely.

The headline is simple. An embedding model sits at No. 1. Not a chatbot, not an image generator, not a giant reasoning model. Google DeepMind's EmbeddingGemma 2 entered the Top 10 straight at the top, and it pushed last week's leader, Cloudflare Clef, down to No. 3. In between sits a 12B model whose only job is to make AI text read like a human wrote it. That tells you a lot about what builders are working on right now.

What happened this week

I pulled the Hugging Face models API on 2026-10-12 at about 16:54 Berlin time with sort=trendingScore. Six of the ten names were not in last week's list. Two of them, Qwen-Image-2.1-Turbo and Underdog Saluki 27B, had model pages created in the last seven days. Four models stayed in the Top 10 from week 41: Cloudflare Clef, the community Qwen-Image-2.1 quantized pack, Lightricks LTX-2.5 and Qwen3.8-27B.

Six names left the list: Laya from Convai Innovations, Aleph Alpha's Kolibri-1, Clef-Flash, the Xing4.0 GGUF, the official Qwen-Image-2.1 and VisionHOPE. That is a lot of churn for a single week, and it shows how short attention spans on the hub really are. A model can lead the chart and be gone seven days later.

One note on method. The community repo abenzerps/Qwen-Image-2.1-Uncensored-Quantized has the same creation date as last week's Uncensored GGUF repo from the same author. It looks like the repo was renamed and now ships more formats, so I treat it as the same entry that held No. 3 last week.

1. google/embeddinggemma-2 (new in the Top 10)

EmbeddingGemma 2 is an open multimodal embedding model from Google DeepMind. It maps text, code, images, video and audio into one shared vector space with 768 dimensions. The whole model has 740M parameters: a 270M text backbone plus a 170M vision encoder and a 300M audio encoder that you only load if you need them.

The model card names the target clearly: phones and laptops. It is meant for on-device search, retrieval augmented generation, classification and clustering. It understands more than 100 languages, claims about 14 percent better results on code tasks than the first EmbeddingGemma, and has an 8K token context, enough for a few minutes of audio or video. It also supports Matryoshka embeddings, so you can cut vectors down to 512, 256 or 128 dimensions and save up to six times the storage.

Snapshot numbers: trendingScore 1,712, 1,868 likes and 115,763 downloads in the 30-day field. The license is Apache 2.0. The page was created on September 14, so this is not a brand-new release. It is a model that took a few weeks to find its audience and then hit the top hard.

2. jialinyyzz/humanizer (new in the Top 10)

humanizer is a 12B model that rewrites AI-written drafts so they read like a person wrote them. It handles emails, essays, reports and forum posts in English and Chinese. The author says it is trained to keep every number, unit, date, name and quote and to add nothing new.

It runs locally, and there are desktop apps for Apple silicon Macs and Windows that work offline. The model card reports that 95 percent of 210 English rewrites were judged human by Originality.ai at its strictest setting. It is also honest about errors: a fact judge flagged 44 of 420 English rewrites, most of them for a single word or phrase, and the author tells users to read the result before sending it.

Snapshot numbers: trendingScore 1,002, 1,259 likes, 46,070 downloads. You can argue about whether a tool like this is a good thing. The demand is obvious, though.

3. Cloudflare/clef (down from No. 1)

Clef was last week's leader. It is a 27B multimodal decision model from Cloudflare, post-trained from Qwen3.8-27B. You hand it a state as text, JSON, images or video plus a schema of typed questions, and it returns a probability for every allowed answer in one forward pass. There is no free text and no output parsing. A small joint schema head on top of the backbone does the scoring.

Clef dropped two places, but it is far from fading. It added 792 likes in a week, going from 1,278 to 2,070, and its trendingScore is 599. Downloads in the 30-day field rose to 15,064. The license is Apache 2.0.

4. abenzerps/Qwen-Image-2.1-Uncensored-Quantized (down from No. 3)

This community repo packs Qwen-Image-2.1 into formats that run on local machines: GGUF, FP8 and INT8 safetensors, NVFP4 and MLX builds for Apple silicon. File sizes go from about 14 GB for BF16 down to 4 GB for the 4-bit versions.

The numbers are big. It has 4,056 likes, up 883 from last week, and more than 2.18 million downloads in the 30-day field, the second highest download count in the whole Top 10. Its trendingScore is 595, just four points behind Clef. The message is clear: people want strong image models they can run at home, without an API and without filters.

5. Qwen/Qwen-Image-2.1-Turbo (new release)

Qwen-Image-2.1-Turbo is the fast version of Qwen's image model. It does text-to-image and image editing in only 8 denoising steps and uses the same 7B visual generation architecture as the full model. The recommended sampling schedule ships with the checkpoint, so it works without manual scheduler setup in Diffusers.

The page went live on October 9, three days before my snapshot, and it already reached trendingScore 578 with 600 likes. Interesting detail: the official full Qwen-Image-2.1 dropped out of the Top 10 this week while the Turbo version came in. Speed wins.

6. Lightricks/LTX-2.5 (biggest riser)

LTX-2.5 is the image-to-video model from Lightricks. Access to the weights is gated on the hub, so you have to accept the terms before you can download it. It slipped one place, from No. 5 to No. 6, but by likes it is the biggest riser of the week.

LTX-2.5 added 1,036 likes in seven days, from 6,363 to 7,399, more than any other model that was in both lists. It also has more than 1.7 million downloads in the 30-day field. The model page dates back to July 23, which makes this one of the most stable names on the chart. Open video generation clearly has a loyal crowd.

7. Qwen/Qwen3.8-27B (holds No. 7)

Qwen3.8-27B is the dense 27B model from the Qwen3.8 family. It is a native vision-language model that understands images and video, with switchable thinking, and the Qwen team pitches it for coding, research and long agentic tasks. It is also the base for several other models on this list: Clef and Underdog Saluki are both built on top of it.

It holds No. 7 with trendingScore 364. It has 17,551 likes, the highest total in the Top 10, and nearly 6.9 million downloads in the 30-day field. That is the gap between lifetime popularity and short-term momentum that makes a trending chart interesting.

8. ConwayResearch/Underdog-Saluki-27B-1.0 (new release)

Underdog Saluki 27B squeezes Qwen3.8-27B into a single 7.89 GB GGUF file, down from 54 GB for the full model. It runs in stock llama.cpp and was tuned to keep tool calling intact, which usually breaks first when you compress a model this hard.

The model card claims 88 of 120 passed tasks on its tool calling benchmark built from the Berkeley Function Calling Leaderboard, against 84 for the full-size model, and 96 percent average retention across nine benchmarks. An optional vision add-on brings image input back. Snapshot numbers: trendingScore 344, 353 likes and 242,329 downloads only four days after the page went live on October 8.

9. LiquidAI/d1-3B (new in the Top 10)

d1-3B is a 3B decision model from Liquid AI, built on LFM2.5-VL-3B. Like Clef, it takes a state of text, JSON or images plus a set of questions and returns calibrated answers in one forward pass, with zero output tokens.

Liquid AI says it is the best decision model under 10B on the Decision Index 0.2.1 with a score of 48.57. It takes 8 ms per decision on an RTX 4090 and 30 ms on an Apple M5 Pro. The suggested uses are routing, triage, moderation, classification and reranking. Snapshot numbers: trendingScore 319 and 327 likes. With Clef and d1 both in the list, decision models now have two seats in the Top 10.

10. canberkkkkkk/ema-lightning (new in the Top 10)

EMA Lightning is the surprise of the week. It is a tiny Turkish text-to-speech model with 8.6M parameters in total, about 34 MB on disk, released under Apache 2.0.

The author reports a word error rate of 0.92 percent on the Freya-TR-Eval set, lower than every other system tested, including ElevenLabs v4 and Gemini 3.8. First audio arrives in under 4 ms on an RTX 4090, it streams, and it runs fully offline on GPU or CPU. Snapshot numbers: trendingScore 317, 374 likes and 20,310 downloads. A one-person project in a single language in the global Top 10 is rare and worth a look.

Why it matters

Look at the list as a whole and a pattern shows up. The top of the chart is about infrastructure, not chat. An embedding model leads. Two decision models answer with probabilities instead of text. Compressed builds like Saluki and the quantized Qwen-Image pack make big models fit on normal hardware. Even the speech model wins on size and speed.

For anyone building with AI, that is the real story. The interesting work right now happens in search, retrieval, routing and running models locally. These are the parts that decide whether an AI product is fast, cheap and private, and the open-source crowd is clearly voting for them.

Dany's take

An embedding model at No. 1 says a lot. Builders are spending their time on search and retrieval right now, not on another chat wrapper. I also like that small models keep winning attention: a 3B decision model, a 34 MB speech model and a 27B model shrunk to under 8 GB are all in the Top 10. Next week I will watch whether EmbeddingGemma 2 holds the top spot and whether humanizer stays this popular.

Data: Hugging Face trending models, pulled on 2026-10-12.