Hugging Face Top 10: DeepSeek V4.1 Flash tops the chart two days after release

Every Monday I pull the public Hugging Face trending ranking, sorted by trendingScore, and write up the Top 10. This is week 2026-W43, the third snapshot in the djcroman series. Last week an embedding model sat at No. 1. This week a brand-new DeepSeek release took the top spot two days after it landed on the hub.

The pattern has flipped hard. Seven of the ten names were not in last week's list. Six of the ten take images as input. Open models are racing on vision right now, not on another chat wrapper. Lightricks LTX-2.5 is still the biggest like gainer among the models that stayed, with 1,044 new likes in seven days, even though it slipped four places on the chart.

What happened this week

I pulled the Hugging Face models API on 2026-10-19 at about 16:34 Berlin time with sort=trendingScore. Three models stayed from week 42: Google EmbeddingGemma 2, the community humanizer rewrite model, and Lightricks LTX-2.5. All three still added hundreds of likes, but each fell four places because the new entries were louder.

Seven names left the list: Cloudflare Clef, the community Qwen-Image-2.1 quantized pack, Qwen-Image-2.1-Turbo, Qwen3.8-27B, Underdog Saluki 27B, Liquid AI's d1-3B decision model, and the tiny Turkish TTS model EMA Lightning. That is almost a full reset of last week's chart. A model can lead one Monday and be gone the next.

Three of the new entries had model pages created inside the last seven days: DeepSeek-V4.1-Flash-1017 on October 17, FLUX 3 Image Klein Comfy on October 14, and FLUX 3 Image Klein on October 13. Liquid AI's LFM2.5-24B-A2B went live on October 12, just outside that seven-day window, so I mark it as new to the Top 10 rather than a brand-new release.

1. deepseek-ai/DeepSeek-V4.1-Flash-1017 (new release)

DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, released on October 17 under the MIT license. DeepSeek says it improves reasoning, coding, agentic work and visual benchmarks, and that it finishes tasks with shorter outputs, which cuts cost. The public safetensors metadata lists about 763B parameters. The repository holds roughly 510 GB across 48 weight shards, stored mostly as FP8 with FP4 experts.

On DeepSeek's own table, the 1017 checkpoint scores 77.0 on DeepSWE v1.1, ahead of Opus 5 and GPT-5.6 Sol in that row, and 42.3 on Terminal-Bench 4.0, up from 31.2 on the previous Flash build. It is tagged image-text-to-text on the hub: it takes images as well as text. The recommended context window is 1M tokens. DeepSeek also opened separate Search and Fetch APIs the same day, but those are API products, not part of the open weights.

Snapshot numbers: trendingScore 1,448, 1,581 likes and 17,924 downloads in the 30-day field. Two days after release, that is enough for a clear No. 1.

2. openbmb/MiniCPM-V-4.7-35B-A3B (new in the Top 10)

MiniCPM-V 4.7 is the latest vision-language model from OpenBMB. The 35B-A3B variant is the first Mixture-of-Experts model in the MiniCPM-V series: 35B total parameters with about 3B active. It keeps the family's single-image, multi-image and video understanding, and adds Lookahead Expert Scheduling so the MoE can run on phones without the usual quality drop.

OpenBMB claims more than 13 tokens per second on an iPhone 17 Pro Max with about 3 GB of memory when the model is quantized and paired with DSpark speculative sampling. The license on the hub is Apache 2.0. The page was created on October 7, so this is a fresh model finding its audience rather than a same-week drop.

Snapshot numbers: trendingScore 850, 1,036 likes and 76,410 downloads. That download count is already the third highest in this Top 10.

3. fishaudio/drama-3 (new in the Top 10)

Drama 3 is Fish Audio's new text-to-speech model for multi-character scenes and long-form audio. It uses a 35B MoE text backbone with about 3B parameters active per token, a codebook multi-token prediction audio decoder with 10 codebooks, and speaker tags for up to 16 characters in one script. Long text is chunked automatically and can stream per chunk.

It launched with day-one support in SGLang Omni and vLLM Omni, which is unusual for a speech model and explains some of the hub attention. The page went live on October 1. Snapshot numbers: trendingScore 709, 736 likes and 11,795 downloads. Speech is rarely this high on the global chart, so Drama 3 is worth watching.

4. black-forest-labs/flux-3-image-klein-comfy (new release)

FLUX 3 Image Klein for ComfyUI is the ComfyUI pack of Black Forest Labs' new 5.6B open image model. One checkpoint handles text-to-image, single-reference edits and multi-reference compositions with up to 10 reference images. The text encoder is Qwen3-VL-4B-Instruct. The 5.6B label is the image transformer only.

This repo ships distilled and base transformers in BF16, int8 and NVFP4, plus the text encoder, prompt upsampler and autoencoder. BFL is clear that short prompts give generic output: the model expects a full scene caption, so the prompt upsampler is part of the default workflow. The page went live on October 14.

Snapshot numbers: trendingScore 698, 891 likes and 393,062 downloads in the 30-day field. That download count is the second highest in the Top 10 and shows how fast ComfyUI packs move once a new open image model lands.

5. google/embeddinggemma-2 (down from No. 1)

EmbeddingGemma 2 led last week's chart. It is Google DeepMind's open multimodal embedding model: text, code, images, video and audio into one 768-dimensional space, about 740M parameters, Apache 2.0, meant for on-device search and retrieval.

It dropped four places to No. 5, but it is not fading. It added 828 likes in a week, from 1,868 to 2,696, the second biggest like gain among models that stayed. Downloads in the 30-day field rose to 167,750. An embedding model that held the top spot for a week and then stayed in the Top 10 is still a strong signal about retrieval demand.

6. jialinyyzz/humanizer (down from No. 2)

humanizer is the 12B local rewrite model that turns AI drafts into text that reads like a person wrote it, in English and Chinese, while keeping numbers, dates, names and quotes. Desktop apps for Apple silicon and Windows run offline.

It fell from No. 2 to No. 6 and still added 713 likes, from 1,259 to 1,972. Downloads jumped to 109,937. Demand for this kind of tool is not going away, even when flashier releases steal the trending score.

7. meta-models/Muse-Spark-Open-816B-A42B (new in the Top 10)

Muse Spark Open is Meta Superintelligence Labs' open weights release of Muse Spark: 816B total parameters with 42B active, a 1M token context, and native multimodality for text, images, audio and video. The model card positions it for coding agents, multi-step tool use and long-horizon agentic work, with extra safety post-training for standalone use.

The license is the Muse Model License, not a standard OSI license, so check the terms before you ship it. The page dates to mid-September; it is new to the Top 10 this week, not a same-week drop. Snapshot numbers: trendingScore 523, 568 likes and 2,109 downloads. The download count is low for its place on the chart, which often means people are starring and reading before they pull half a trillion parameters.

8. black-forest-labs/flux-3-image-klein (new release)

FLUX 3 Image Klein is the official distilled checkpoint of the same 5.6B model, without the ComfyUI packaging. Same capabilities: prompt plus up to 10 references, text-to-image and editing through one transformer. For fine-tuning and LoRA work, BFL points to the separate Base repo.

The page went live on October 13, one day before the Comfy pack. Snapshot numbers: trendingScore 474, 517 likes and 62,006 downloads. Having both the official weights and the Comfy pack in the Top 10 in the same week is unusual and underlines how quickly the local image community moves when a new open model ships.

9. LiquidAI/LFM2.5-24B-A2B (new in the Top 10)

LFM2.5-24B-A2B is Liquid AI's largest LFM2.5 model: 24B total parameters with only 2B active, built for fast private agents on-device. Liquid AI says it scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, runs up to 1.7 times faster than comparable MoEs on-device, and reaches about 184 tokens per second on an Apple M5 Max.

The page was created on October 12 at 05:35 UTC, just outside my seven-day "new release" window, so I list it as new to the Top 10. Snapshot numbers: trendingScore 471, 487 likes and 14,151 downloads. Last week Liquid AI's smaller d1-3B decision model was in the list. This week the flagship reasoning MoE takes that seat instead.

10. Lightricks/LTX-2.5 (biggest riser)

LTX-2.5 is Lightricks' gated image-to-video model. It fell from No. 6 to No. 10 on trendingScore, but by likes it is the biggest riser of the week among models that were in both Top 10s.

It added 1,044 likes in seven days, from 7,399 to 8,443, and the 30-day download field sits above 2.0 million. The model page dates back to July 23. Open video generation still has the most loyal crowd on this chart, even when brand-new multimodal releases steal the top ranks.

Why it matters

Look at the list as a whole and the story is vision plus efficiency. Three image-text-to-text models, two image-to-image packs from the same FLUX 3 launch, one image-to-video veteran, one multimodal embedding model, one multi-speaker TTS model, and two text models that are really about local rewrite and on-device agents. Chat for its own sake is barely on the board.

Seven new names in one week also show how short the hub's attention span is. Last week's decision models and compressed GGUF packs are gone. This week's winners are either brand-new open weights with hard benchmark claims, or packaging that makes those weights easy to run in ComfyUI and on phones.

Dany's take

DeepSeek V4.1 Flash taking No. 1 two days after release is the headline, but the deeper signal is the vision count. Six of ten models take images as input. Open models are racing on vision now, not on chat. I also keep watching LTX-2.5: it can fall four places on trendingScore and still add more than a thousand likes. Next week I will watch whether DeepSeek holds the top and whether both FLUX 3 packs stay in the list together.

Data: Hugging Face trending models, pulled on 2026-10-19.