Hugging Face Top 10: DeepSeek V4.1 Flash stays on top as tiny LFM3 closes in

Every Monday I pull the public Hugging Face trending ranking, sorted by trendingScore, and write up the Top 10. This is week 2026-W45, the fifth snapshot in the djcroman series. DeepSeek V4.1 Flash 1017 is at No. 1 for the third week in a row. But this time the lead is tiny: one single trending point.

The model right behind it is the opposite of a 763B giant. Liquid AI's LFM3-1.7B has 1.73B parameters and, according to its card, needs about 1 GB of memory at 4-bit precision. Behind those two, the list is the most mixed it has been since I started: speech recognition, document OCR, a trillion parameter Mistral model, 3D generation, video with sound and a small 3D coding model built by one person.

What happened this week

I pulled the Hugging Face models API on 2026-11-02 at 16:28 Berlin time with sort=trendingScore. Only three models stayed from week 44: DeepSeek V4.1 Flash 1017 at No. 1, Lightricks LTX-2.5 at No. 6, the same place as last week, and Nanbeige 4.3 33B, which fell from No. 3 to No. 10.

Seven names left the list. Recursive Flint dropped from No. 2 to No. 16, Spark-X2.6-4B to No. 20, the FLUX 3 Image Klein pack for ComfyUI to No. 13, Qwen3.8-27B to No. 12, the Qwen-Image-2.1 Uncensored Quantized pack to No. 11 and MiniMax M3.1 Flash all the way to No. 49. MOSS-Image-SFT is no longer in the top 50 at all. That makes it the third week in a row with seven changes out of ten.

Only two of the seven newcomers had their model page created inside my seven day window: Idol from Speridlabs on October 28 and Pika H3 Flex on October 29. The others are older pages that caught a wave this week. Baidu's Unlimited OCR 2 page went live on October 26 at 03:57 UTC, about half a day before my window starts, so I list it as new in the Top 10 rather than as a new release. Mistral Large 4 dates back to September 30 and NVIDIA's Nemotron 3.5 Transcribe to October 7. Neither of those two, nor LFM3, was in the top 50 of the trending list a week ago.

A note on the numbers. trendingScore is the Hugging Face ranking signal and reflects recent activity. Likes are all time totals. The downloads field from the API covers the last 30 days. That is why a model at No. 2 can have fewer likes than the model at No. 4.

1. deepseek-ai/DeepSeek-V4.1-Flash-1017 (holds No. 1)

DeepSeek-V4.1-Flash-1017 is the October 17 update to DeepSeek-V4.1-Flash, released under the MIT license. DeepSeek says it improves reasoning, coding, agentic tasks and visual benchmarks while finishing tasks with shorter outputs. The safetensors metadata on the hub lists about 763B parameters and the page is tagged image-text-to-text.

On DeepSeek's own table the 1017 checkpoint scores 42.3 on Terminal-Bench 4.0, up from 31.2 on the previous Flash build. The recommended context window is 1M tokens with an output budget of at least 256K tokens. The release still ships a Python reference encoder instead of a Jinja chat template, plus a separate set of Rust libraries called deepseek-recipe for production prompt formatting.

Snapshot numbers: trendingScore 687, 4,626 likes and 177,009 downloads in the 30 day field. It added 1,517 likes in seven days, again the biggest gain among the models that stayed, and its 30 day downloads almost doubled from 95,889. Its trendingScore cooled from 1,090 to 687, which is why LFM3 could get this close.

2. LiquidAI/LFM3-1.7B (new in the Top 10)

LFM3-1.7B is Liquid AI's text model for local agents and specialized applications. It uses what Liquid calls a tri-hybrid architecture: 24 layers, of which 13 are gated short convolutions, 5 use sliding window attention and 6 use global attention. The model has 1.73B parameters, a 256K token context window and a 128,000 token vocabulary, and it covers 16 languages. Liquid says it was post-trained for agent harnesses such as Hermes Agent, Pi and OpenClaw.

The edge pitch is concrete. At 4-bit precision it takes about 1 GB of memory, and the repo offers BF16, FP8, NVFP4, GGUF, MLX and ONNX variants. It supports deferred tools, so a harness can keep most tool definitions out of the prompt until a search tool returns them. Liquid also ships fine-tuning recipes for SFT, DPO, GRPO and continued pretraining. The license is the LFM Open License v1.0, not a standard OSI license.

On Liquid's own table, LFM3-1.7B scores 71.39 on BFCL V4 against 62.83 for Qwen3.5-9B, and its average over all listed benchmarks is 49.37 against 47.24 for Qwen3.5-9B. I like that the card is honest about limits: on SWE-bench Verified it trails Qwen3.5-9B with 43.40 versus 51.20, and Liquid itself says the model is not recommended for agentic coding or complex multi-step reasoning.

Snapshot numbers: trendingScore 686, 785 likes and 42,384 downloads. The page was created on October 23.

3. nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B (new in the Top 10)

Nemotron 3.5 Transcribe is NVIDIA's open weight, offline speech and audio model. It transcribes recordings in 40 languages, handles audio up to one hour long and has a 65,536 token context. Beyond plain transcription it does speech translation, speaker attributed transcription, word timestamps, language identification and audio question answering, and you pick the transcription style with a text instruction.

The architecture is a Mamba-2 and Transformer hybrid mixture of experts with an audio encoder and a speaker encoder. NVIDIA lists 34B total parameters with about 4.3B active, including the audio encoder, and says it fits on one H100 80 GB. FP8 and NVFP4 variants are available. The license is OpenMDW-1.1.

On NVIDIA's results table it reaches a mean word error rate of 4.71 on nine public English sets of the Open ASR Leaderboard, against 4.94 for Qwen3-ASR-1.7B, the next best open weight model in that comparison. The card also says the fastest serving numbers come from the NIM implementation, which is not yet part of the open source release.

Snapshot numbers: trendingScore 583, 653 likes and 32,439 downloads. The page dates from October 7.

4. mistralai/Mistral-Large-4-1T-A52B (new in the Top 10)

Mistral Large 4, nicknamed Le Chonk on its own card, is Mistral's largest model so far. It is a natively multimodal mixture of experts with about 1 trillion total parameters and 49B active per token, or 52B when you count embeddings and output layers. It uses 128 routed experts, four of them active per token, plus one shared expert. Input is text and images, output is text, and the context window is 1M tokens.

Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers in Europe. It has three reasoning efforts, low, medium and high, and native function calling with strict JSON schema output. On the Artificial Analysis Coding Agent Index, Mistral reports 48 percent, which it calls effectively tied with GLM-5.3 while using 47 percent fewer tokens per task. On the Artificial Analysis Intelligence Index v5 it scores 26, and the card openly says the leading Chinese open weight models score higher there.

The license is the Mistral Ecosystem Open-Weight License, a modified MIT license. Running a model as a service business needs a separate commercial license. Snapshot numbers: trendingScore 491, 1,479 likes and 5,500 downloads. Few people download a 1T model, but a lot of people like it.

5. baidu/Unlimited-OCR-2 (new in the Top 10)

Unlimited OCR 2 is Baidu's document parsing model, released on October 27 according to its card, with a paper on arXiv dated November 2. The core idea is Sparse Reference Sliding Window Attention, SR-SWA. It squeezes the interaction between text and page images down to a fixed set of working references, for example 512 tokens, so latency and memory per decoding step stay almost constant no matter how long the document is.

Baidu claims state of the art results on long document benchmarks, from dozens of pages up to hundred page documents, and released a new benchmark called MarathonDocBench for that. You can parse all pages of a document in one request so the model sees cross page context. It runs with Transformers for quick tests and a customized SGLang build for serving. The safetensors metadata lists about 3.3B parameters and the license is MIT.

Snapshot numbers: trendingScore 465, 550 likes and 147,588 downloads in the 30 day field. That is the third highest download count in this Top 10, behind only LTX-2.5 and DeepSeek. People clearly have PDFs to process.

6. Lightricks/LTX-2.5 (holds No. 6)

LTX-2.5 is Lightricks' gated video model. The hub tags list image to video, text to video, video to video and audio and video combinations, and the page goes back to July 23, which makes it the oldest model in this week's list again. Because access is gated I cannot read the full card without logging in, so I stick to what the public API shows.

It stays at No. 6 and keeps growing. Likes rose from 9,395 to 10,310, a gain of 915, the second biggest among the three models that stayed. The 30 day download field went from 2,117,147 to 2,467,613, the highest in this Top 10 by far. trendingScore is 378. LTX-2.5 is the only model that has made all five of my weekly Top 10 snapshots so far, from No. 5 in week 41 to No. 6 now.

7. ccore/Qwen3.5-4B-3D-Coder (new in the Top 10)

Qwen3.5-4B-3D-Coder is the most charming entry this week. It is an independent fine-tune of Qwen3.5-4B by the user CCORE, explicitly not an official Qwen release, and it is labeled an early preview. The goal is to test whether a small model can do general coding and especially generate interactive 3D scenes with Three.js.

The author has been shipping in public: an October 27 checkpoint update with better one shot success, full FP16 weights, a technical report and the training dataset on October 29, and optional vision support on October 30. With the original Qwen3.5-4B projector, about 644 MiB, the GGUF versions can take a reference image and turn it into a Three.js scene. The dataset has 4,146 scene examples under MIT, and the model itself is Apache 2.0. The page has no pipeline tag, which is why it shows up as untagged in my chart.

Snapshot numbers: trendingScore 375, 535 likes and 63,499 downloads. The page was created on October 25.

8. speridlabs/idol (new release)

Idol turns a single picture of an object into a 3D Gaussian splat, including the sides the picture never showed. You give it a render, a generated image or a photo. It removes the background, frames the object and generates up to 1,048,576 Gaussians, written out as a standard 3DGS .ply file.

The card says it runs on consumer GPUs with 7.2 GB of peak VRAM and can run in under 2 seconds. A full request uses about 1.80B parameters across its parts: a 370.6M denoiser, octree and Gaussian decoders, BiRefNet for background removal, a DINOv3 encoder and the FLUX.2 VAE. One catch: two of those encoders sit behind gated Hugging Face repos, so you have to accept their terms before Idol runs. In Speridlabs' own benchmark against TripoSplat, TRELLIS and LGM, Idol leads most quality columns at 6.8 seconds per request with 50 steps and 262,144 Gaussians. The license is Apache 2.0.

Snapshot numbers: trendingScore 358, 376 likes and 7,787 downloads. The page went live on October 28.

9. Pika-Labs/Pika-H3-Flex (new release)

Pika H3 Flex is Pika's post-trained video and audio model, built as adapters on top of MiniMax-H3. It generates video with synchronized stereo audio in six sampling steps. There are two task adapters: one for text and first or last frame conditioning, and one for multimodal references with up to 9 images, 3 videos and 3 audio clips, 12 files in total.

Local inference produces native 480p or 768p clips of 5 to 15 seconds at 24 fps, with 32 kHz stereo audio, across 11 aspect ratios from 21:9 to 9:21. Higher resolutions up to 4K, audio stems and prompt enhancement are only available through Pika's hosted API. The safetensors metadata lists about 33.1B parameters and the license is the MiniMax H3 community license.

Snapshot numbers: trendingScore 326, 342 likes and 491 downloads. The page was created on October 29. Interesting side note: the ComfyUI repack of its base model, Comfy-Org/MiniMax-H3, is among the five most downloaded models on the whole hub in the 30 day field.

10. Nanbeige/Nanbeige4.3-33B (down from No. 3)

Nanbeige4.3-33B is the flagship of the Nanbeige 4.3 series and is pitched as a teacher model: build a strong agent first, then distill its skills into smaller models. It targets complex coding and what the card calls professional cowork tasks, in English and Chinese, under Apache 2.0, with up to 524,288 tokens of context.

Last week I praised the card for showing rows where it loses. Nanbeige reports 65.5 on DeepSWE against 42.5 for Qwen3.8-27B, but also shows Qwen3.8-27B slightly ahead on SWE-Bench Pro with 61.7 versus 60.3.

It fell seven places, but usage grew. Likes went from 536 to 997, a gain of 461, and the 30 day download field jumped from 2,787 to 16,976. trendingScore is 320. A drop in the ranking after launch week is normal; the downloads say people are actually trying it.

Why it matters

Line up the list by job and there is no single theme this week, and that is the story. Three models are tagged image-text-to-text, but they do very different things: a general agent model from DeepSeek, a trillion parameter generalist from Mistral and a document parser from Baidu. Two are text generation models, two make video, one transcribes speech, one builds 3D objects and one writes 3D scenes in code.

Two patterns still stand out. First, size is spreading in both directions. The list goes from 1.73B parameters for LFM3 to about 1 trillion for Mistral Large 4, and the small one is only one trending point behind the No. 1. Second, active parameters stay small even at the top. Mistral activates about 52B of 1T, Nemotron Transcribe about 4.3B of 34B. The industry keeps building huge models that run like medium ones.

Also notable: two of the newcomers are not from big labs. CCORE is a single developer fine-tuning a 4B model in public, and Speridlabs is a small team. The Hugging Face trending list still rewards useful, well documented releases, not just famous logos.

Dany's take

DeepSeek holds No. 1 for a third week, but only by a single trending point over Liquid AI's 1.7B LFM3, which fits in about 1 GB at 4-bit: the next fight is about agents on your phone, not in the data center. I also like how many cards this week say clearly what their model is bad at. Liquid says LFM3 is not for agentic coding, Mistral admits the Chinese models score higher on the big index, and that makes the rest of their claims easier to trust. Next week I will watch whether LFM3 actually takes the top spot and whether the long tail of 3D and video models stays in the list.

Data: Hugging Face trending models, pulled on 2026-11-02 at 16:28 Berlin time. Model details from each linked model card on hf.co.