Top 5 Hugging Face Models This Week

DeepSeek updated its flagship model again, this time with the index 1017, and the authors call it a minor step forward in reasoning, coding and agentic tasks. The model weighs 763 billion parameters, and even the compressed version needs about 460 GB, so installing it at home is not possible. Downloads jumped almost tenfold this week, from about 10 thousand to over 105 thousand. I work with agents in Claude Code every day, and with Chinese models of this size I keep seeing the same pattern: strong numbers in the tables, weaker once you give them real agentic work.

💻 Server-grade hardware only — you cannot run this at home. Even the compressed version needs about 460 GB.

The Spark team released Spark-X2.6-4B, a compact model built for on-device use, a little over a month after the previous version. It has only 4.1 billion parameters, holds a context window of close to a million tokens, and ships under Apache 2.0, so commercial use is free. The architecture stayed the same as the previous version, and the authors say all the gains came from extra post-training, with the model matching or beating some 9-billion models on their benchmarks. I like that it has three thinking levels, so for simple tasks you can turn reasoning off and not pay for it in time.

💻 A regular laptop with 8 GB of RAM or more — the model is tiny, 4.1 billion parameters.

Black Forest Labs describe FLUX 3 Image Klein as an open-weights, 5.6-billion-parameter model for generating and editing images, and this time they packaged it for ComfyUI themselves with ready-made workflows, so there is nothing to assemble by hand. It takes up to ten reference images and combines text-to-image generation, single-photo editing and multi-photo composition in one model, instead of three separate ones. The lighter version runs on a GPU with 8 GB of VRAM, which fits even an older RTX 3070. Downloads more than doubled this week, from about 338 thousand to 750 thousand, so interest is clearly not fading.

💻 A GPU with 8 GB of VRAM or more — heavier versions need a stronger card.

Lightricks describe LTX-2.5 as an open-weights model built for local execution and fine-tuning, and its main job is generating synchronized video and audio from a text, image or video input. A clip can run up to 121 frames and now be assembled from several connected scenes while keeping the same character — before, that had to be stitched together by hand. The model comes as two 22-billion-parameter versions, a fast distilled one and a full one for fine-tuning, but the authors never name an exact GPU for running it, which is an annoying gap if you plan to install it at home. Under the license, commercial use is free as long as your company's revenue is under 10 million dollars a year.

💻 The authors give no exact number. Going by the model's size (22 billion parameters), that is roughly a GPU with 13 GB or more for the compressed version, a stronger GPU or a Mac with more memory for the full one.

Qwen released Qwen3.8-27B, a 27.8-billion-parameter model that understands not just text but images and video too, and holds a context window of up to a million tokens. On the OSWorld computer-use benchmark it scored 84.3, and on GPQA Diamond 89.2. The authors call it a step forward in coding, professional work and long agentic tasks. I like that a model like this can run on your own GPU instead of only through someone else's API — for agentic work that means more control over what happens to your data.

💻 A Mac with 32 GB of memory or a GPU with 24 GB of VRAM.