Hugging Face Top 10 of the week: Cloudflare Clef at No. 1

Every Monday I pull Hugging Face's public trending ranking, sorted by trendingScore, and write up the Top 10. This is week 2026-W41, the first snapshot in the djcroman series, so it also sets the baseline for next week's risers and fallers.

The chart is not a pure downloads leaderboard. Hugging Face mixes engagement into a trending score, and the models that move fastest often sit far from the ones with the largest 30-day download counters. The weekly roundup is meant to catch what the open-source crowd is circling right now: decision models that never generate free text, GGUF packs that local runners can actually load, European labs shipping MoE reasoning weights, and video stacks that refuse to leave the board.

What happened

On 2026-10-05 at about 10:29 Europe/Berlin I pulled the public Hugging Face models API with sort=trendingScore. Four of the Top 10 had model pages created in the previous seven days: Cloudflare/clef, Aleph-Alpha/Kolibri-1, Cloudflare/clef-flash, and PSRben/VisionHOPE. That is a heavy NEW week for a baseline snapshot.

The story of the board is not one architecture. Image-text-to-text decision and multimodal models take three slots. Text-to-image takes two. Text generation takes two. Text classification, image-to-video, and image classification each take one. Cloudflare alone occupies No. 1 and No. 6 with the Clef family. Qwen occupies No. 7 and No. 9 with an official multimodal line and an official image model, while a community uncensored GGUF of that same image line sits at No. 3.

Likes and trending score diverge hard. Qwen/Qwen3.8-27B carries 16,958 likes, the largest total in the Top 10, yet ranks seventh on momentum. Cloudflare/clef has 1,278 likes and still leads. That is the point of a trending lens: early engagement can outrank lifetime popularity.

1. Cloudflare/clef (NEW)

Cloudflare/clef is a 27B multimodal decision model from Cloudflare. It does not generate free-form text. You give it a state (text, JSON, images, or video) plus a schema of typed questions, and it returns a probability for every allowed option in a single forward pass. The model card says it is post-trained from Qwen/Qwen3.8-27B, ships a joint schema head on top of the backbone, and is released under Apache-2.0.

Pipeline tag on the hub: image-text-to-text. Snapshot numbers: trendingScore 1,188, 1,278 likes, 4,214 downloads (30-day field), created 2026-09-30. It hit No. 1 about five days after the page went live. The card also documents Decision Index benchmarks and a SystemOne / Jev compatible API surface, which helps explain why a brand-new decision model can lead a public trending chart so quickly.

2. convaiinnovations/laya

convaiinnovations/laya is Laya, a multilingual non-autoregressive System 1 decision model from Convai Innovations. Like Clef, it answers typed questions with calibrated probabilities and never generates text to parse. The hub repo is the family hub: English root (~421M params on ModernBERT-large), plus multilingual and typed-decisions checkpoints. License on the card: Apache 2.0.

Pipeline tag: text-classification. Snapshot: trendingScore 754, 5,184 likes, 3,752 downloads (30-day), created 2026-09-18. It is not NEW this week. The rank reads as sustained attention for a narrow, useful decision runtime rather than a launch spike. The card positions Laya as a fast self-hosted alternative for triage, routing, and schema-driven decisions across 100+ languages.

3. abenzerps/Qwen-Image-2.1-Uncensored-GGUF

abenzerps/Qwen-Image-2.1-Uncensored-GGUF is a community GGUF and companion-weight pack of Qwen/Qwen-Image-2.1 for local image generation, mainly aimed at ComfyUI with ComfyUI-GGUF. The card lists multiple quantizations (for example Q4_K_M at about 4.6 GB) plus text encoder and VAE files, names the source revision, and states the license as Qwen Research License. The page also states that this release has no built-in safety checker or content filter.

Pipeline tag: text-to-image. Snapshot: trendingScore 703, 3,173 likes, 1,553,744 downloads (30-day), created 2026-09-20. I leave the fork in the ranking because that is what the API returned. Readers can weigh community packaging against the official line at No. 9. The download counter alone shows how hard local runners chase quantized image stacks.

4. Aleph-Alpha/Kolibri-1 (NEW)

Aleph-Alpha/Kolibri-1 is Kolibri 1 from Aleph Alpha in Germany: a mixture-of-experts reasoning model focused on German and English. The model card lists about 78B total parameters with about 3.46B active per token, native context of 262,144 tokens (validated up to 1,048,576), explicit reasoning mode, tool calling, and Apache 2.0 weights. Release date on the card: 3 October 2026. Knowledge cutoff listed as 18 June 2026 for both EN and DE.

Pipeline tag: text-generation. Snapshot: trendingScore 457, 467 likes, 1,135 downloads (30-day), created 2026-10-02. NEW and already inside the Top 5. European labs rarely punch this high this fast on the global HF chart. The card is unusually detailed on training mix, evals, and serving via aleph-alpha-inference / vLLM, which matches a lab trying to make a bilingual MoE usable in production rather than only as a demo.

5. Lightricks/LTX-2.5

Lightricks/LTX-2.5 is LTX-2.5, an open-weights video, audio, and world-simulation stack from Lightricks. The card describes synchronized high-fidelity video and audio from text, image, and video inputs, native multishot generation, a custom Gemma 4 12B text encoder, and split Comfy-aligned checkpoints (including a 22B distilled transformer). Licensing is under the LTX-2.x Community License, with commercial terms that depend on annual revenue.

Pipeline tag: image-to-video. Snapshot: trendingScore 455, 6,363 likes, 1,626,951 downloads (30-day), created 2026-07-23. Not NEW. It holds one of the highest like totals on this board without needing a fresh launch spike. Video generation stays sticky even when language models dominate the broader news cycle.

6. Cloudflare/clef-flash (NEW)

Cloudflare/clef-flash is the lighter Clef sibling: a 9B multimodal decision model post-trained from Qwen/Qwen3.5-9B, same typed-question probability API, Apache-2.0. The Decision Index table on the card shows much lower median latency than full Clef (about 38.8 ms versus about 209.3 ms in Cloudflare's published run), which is the product pitch for flash variants.

Pipeline tag: image-text-to-text. Snapshot: trendingScore 431, 450 likes, 6,372 downloads (30-day), created 2026-09-30 (one second after full Clef in the API timestamps). Cloudflare took two of the Top 6 with one product family in the same release window. That is the clearest product headline on the board after Clef's No. 1.

7. Qwen/Qwen3.8-27B

Qwen/Qwen3.8-27B is Qwen's 27B dense multimodal model: a causal language model with a vision encoder, native image and video understanding, flexible thinking control, and a native context of 262,144 tokens (extensible toward 1M with YaRN-style scaling). The card positions it for coding, professional work, research, and long-horizon agentic tasks.

Pipeline tag: image-text-to-text. Snapshot: trendingScore 343, 16,958 likes, 6,821,761 downloads (30-day), created 2026-08-05. Largest like count in this Top 10, and still only seventh on momentum. That gap is useful: permanent weight on the hub is not the same as this week's velocity. Clef can outrank it on trendingScore even with fewer than one tenth of the likes.

8. Venastine-Research/Xing4.0-29B-A4B-GGUF

Venastine-Research/Xing4.0-29B-A4B-GGUF packages Xing4.0-29B-A4B, a MoE language model from China Telecom Artificial Intelligence Technology (Xing series, formerly TeleChat). The card describes 29B total parameters with 4B active per token, 256K context extensible to 512K, agent-oriented architecture, and Ascend NPU / MindSpore training claims. This hub entry is the GGUF packaging line for local and compatible inference stacks.

Pipeline tag: text-generation. Snapshot: trendingScore 322, 343 likes, 14,361 downloads (30-day), created 2026-09-17. Modest likes, still inside the Top 10. Quantized packs keep showing up because people who run models on their own hardware can actually load them. I am not restating a license here beyond what was clear on the fetched card text.

9. Qwen/Qwen-Image-2.1

Qwen/Qwen-Image-2.1 is the official Qwen text-to-image and image-editing model. The card highlights a compact 7B visual generation component, native transparent (RGBA) generation, editing with up to 10 reference images, and licensing under the Qwen Research License Agreement.

Pipeline tag: text-to-image. Snapshot: trendingScore 299, 2,956 likes, 90,003 downloads (30-day), created 2026-09-14. It sits two places under its uncensored community GGUF fork. Packaging and distribution format matter as much as the base weights when the weekly metric is momentum among people who download and run.

10. PSRben/VisionHOPE (NEW)

PSRben/VisionHOPE publishes official pretrained weights for VisionHOPE, described on the card as visual backbones framed as self-modifying learning systems. The repo includes hierarchical T/S/B checkpoints for ImageNet-1K classification, COCO detection and instance segmentation, and ADE20K semantic segmentation, with a linked paper (arXiv 2609.33325) and code.

Pipeline tag: image-classification. Snapshot: trendingScore 287, 408 likes, 1,516 downloads (30-day), created 2026-09-29. NEW, and it closes the Top 10. Vision models still find room beside language and video heavyweights when trending score lines up. I leave license details to the linked paper and code repo where the fetched card did not spell them out clearly.

How to read the numbers

This week is the baseline: no prior weekly snapshot yet, so no risers versus last week. Those markers start next Monday. NEW means the model page was created in the last seven days relative to the pull. Likes are cumulative. The downloads field on the Hugging Face API remains a rolling 30-day count, which is why the cards and this article lead with trending rank and likes rather than treating downloads as a weekly score.

TrendingScore can elevate a brand-new model with sharp early engagement over an older model with more absolute likes. That is why Clef can outrank Qwen3.8-27B. The score is a momentum lens, not a lifetime trophy case. It is also why a GGUF fork can sit above an official line: the people loading weights this week may prefer the pack that fits their VRAM and workflow.

Why it matters

Open-source model discovery is noisy. A fixed Top 10 from the same public API gives builders and curious readers one place to see which architectures and labs are moving the hub right now, week after week, with comparable methodology.

The board also shows a shift in what "model" means on Hugging Face. Decision models that output probabilities, image-to-video stacks, GGUF packs, bilingual MoE reasoning weights, and classic multimodal generators share the same Top 10. Kolibri-1 is the European headline. Cloudflare's double Clef entry is the product headline. The GGUF rows explain why quantized forks keep beating official cards on weekly momentum. Laya holding No. 2 shows that narrower task models still earn attention when they are useful and fast.

Dany's take

Cloudflare's Clef hitting No. 1 five days after release is the cleanest signal on the board. Decision models that output probabilities, not text, can lead the public trending chart. Clef-flash in the same window makes it a family story, not a one-off upload.

Laya at No. 2 is the reminder that small, specialized decision runtimes still matter when latency and calibration are the product. Kolibri-1 at No. 4 as a NEW release puts Aleph Alpha on the weekly map with a serious German-English MoE card. LTX-2.5 keeps video generation in the conversation without needing a fresh launch spike. Qwen's official lines still dominate likes and long-term downloads even when community packaging wins a weekly slot.

I am leaving the uncensored Qwen-Image fork in place because the ranking is what the API returned. Readers can weigh forks versus official lines for themselves; the official Qwen-Image 2.1 is still on the same list at No. 9.

Next Monday the same pull gets a second data point. That is when risers and fallers start, and when a one-week wonder has to prove it can hold the slot.

Sources

• Hugging Face models sorted by trending: huggingface.co/models?sort=trending
• Cloudflare/clef
• convaiinnovations/laya
• abenzerps/Qwen-Image-2.1-Uncensored-GGUF
• Aleph-Alpha/Kolibri-1
• Lightricks/LTX-2.5
• Cloudflare/clef-flash
• Qwen/Qwen3.8-27B
• Venastine-Research/Xing4.0-29B-A4B-GGUF
• Qwen/Qwen-Image-2.1
• PSRben/VisionHOPE

Snapshot pulled from the public Hugging Face API on 2026-10-05 at 10:29 Europe/Berlin.