Top 5 Hugging Face Models This Week

DeepSeek pushed out an updated version of its flagship model, V4.1 Flash, now carrying the index 1017, and the authors call it an improvement for reasoning, coding and agentic tasks. It weighs 763 billion parameters, and even a compressed version for local use takes several hundred gigabytes, so running it on a home computer is not realistic. On one of the autonomous coding benchmarks it scored 77 points, which does look strong on paper. I work with agents in Claude Code every day, and with Chinese models of this size the pattern usually repeats: good on paper, 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.

OpenBMB showed that their new model for understanding images and video, MiniCPM-V 4.7, runs right on an iPhone, with no server and no internet connection — the authors claim over 13 tokens per second on an iPhone 17 Pro Max, using about 3 GB on the device. The model reads a single image, several images together, or video, so you can hand it a screenshot or a clip and get a plain description back. For a desktop there is a GGUF version too, so setting it up on a regular computer is not a problem either. It is a neat technical trick, but I would not rely on it for serious work — this class of model is fine for simple, repeated tasks, not for real agentic work.

💻 A Mac with 64 GB of memory or a 48 GB GPU for desktop use. On phone, the authors say about 3 GB is enough.

FishAudio released the third generation of its voice model, Drama 3, and they call it their new text-to-speech model built for multi-character scenes and long-form content. In one recording it can give up to 16 different voices, clone a voice from a short sample, and control tone through simple text tags, marking a line as angry or whispered, for example. The authors mention 83 supported languages and automatic splitting of long texts, so narrating a whole book chapter does not mean cutting the text by hand yourself. On heavy server GPUs they get under 100 milliseconds per line, but what a regular graphics card would give is not stated in the card.

💻 Estimated: a Mac with 64 GB of memory or a 48 GB GPU; the authors do not give exact requirements for a regular computer.

Black Forest Labs describe FLUX 3 Image Klein as an open-weights model with 5.6 billion parameters for generating and editing images, and this time they packaged it for ComfyUI themselves, with ready-made workflows included, so there is no need to assemble the model from pieces yourself. You can generate an image from a description, edit a single photo by instruction, or compose a scene from several reference images. The lighter version runs on 8 GB of VRAM, which fits even an older RTX 3070. The weights are open, and under the license, commercial use of the output is free as long as your company's revenue is under 5 million dollars a year.

💻 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, with its established use being generating synchronized video and audio from a text, image, or video input. It comes as two 22-billion-parameter transformer versions, a fast distilled one and a full one for fine-tuning. For weaker GPUs the authors mention a lighter mode with memory offload, but they do not name an exact minimum card, which is an annoying gap if you are planning to run 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, only a lighter mode for weaker cards. Going by the model's size (22 billion parameters), that is roughly a GPU with 13 GB or more in the compressed version, a stronger Mac or GPU for the full one.