GitHub radar
Top 5 Hugging Face Models This Week
This week on Hugging Face covered three very different things at once: turning a single photo into a 3D model, a video model built to run and be fine-tuned at home, and a 2.8-trillion-parameter cloud model not worth even trying on a home machine. The spread in size and purpose is wider than usual, and half the list are models that actually run on an ordinary computer.
Idol turns a single photo of an object into a full 3D model: the program separates the subject from the background on its own and fills in the sides the camera never saw. By the author's own benchmark, the result is more accurate than three other open tools built for the same job, TripoSplat, TRELLIS and LGM. The model itself is small, but it is stitched together from several parts, and two of them are gated Hugging Face models that require a separate access request tied to an account. It comes from a single developer working under the name speridlabs, not a large lab, which is unusual for something that climbed this fast in one week.
Why a vibe-coder should care
For anyone building 3D scenes, props or avatars for games and visuals, this is a way to get a usable volumetric model from one photo without a 3D scanning studio or a paid service. The catch is the extra step of requesting access to two gated models before the first run.
A GPU is required — the author says a regular gaming card handles one object in about 7 seconds
Open on Hugging FaceKimi K3.1 Pro is Moonshot AI's new flagship model, able to read both text and images and built specifically to work inside AI agents: coding, research, financial analysis. The company calls it the most capable Kimi model to date and claims a real gain over the previous K3 in coding and factual accuracy. It weighs close to 2.8 trillion parameters, so there is no home machine for it; it is meant for servers and reaches people through Moonshot's own cloud API. The weights are published openly, but under the company's own license rather than Apache or MIT, so a commercial product built on it needs to check those terms first.
Why a vibe-coder should care
For anyone plugging an external model into their own agent through an API, this is a fresh and, by the company's own claim, strong option, but the license is custom rather than standard-open, and there is no free local run.
Server-grade hardware — will not run at home
Open on Hugging FaceRipple-65B Preview is a new language model from a small company, Ricursive Intelligence, and by its own tests it is particularly strong at coding and agent-style tasks. It has 65 billion parameters against the 500 to 975 billion of the rivals it is aiming at, and the company claims comparable quality despite the much smaller size. Context runs up to a million tokens, enough to hold a sizeable code project in one go. The license is Apache 2.0, so it can go into any project, commercial included, without asking permission.
Why a vibe-coder should care
For anyone who wants flagship-level help with writing and fixing code without renting a rack of GPUs, this is one of the few models at this level of ambition that can actually sit on a Mac or a single home workstation.
A Mac with 64 GB of memory or more, or a GPU with 48 GB
Open on Hugging FaceLTX-2.5 is an open video model from Lightricks, and the authors say plainly it is meant to run and be fine-tuned on someone's own hardware, not only rented through the cloud. From text, a photo or another video it builds not just one clip but several connected shots in a row, keeping the same face, voice and setting across the cuts. A new video decoder, according to the developers, gave sharper faces and textures and more natural motion than the previous version. The model weighs 22 billion parameters, and the Community license is free for commercial use for companies under 10 million dollars in annual revenue; past that line a paid agreement is required.
Why a vibe-coder should care
For anyone assembling video clips or ad cuts that need the same character across several scenes, this is a rare open model that can be installed and fine-tuned locally instead of rented as a subscription, though a large business will eventually have to pay for it.
A Mac with 32 GB of memory or more, or a GPU with 24 GB
Open on Hugging FaceQwen3.5-4B-3D-Coder is a fine-tune of Qwen3.5 from an independent developer group, CCORE, and the authors themselves call it an experiment: can a small model write ordinary code and also build good-looking 3D scenes. From a text description it produces a ready, interactive scene that opens straight in a browser, and since the end of October it builds the same kind of scene from a photo as well. It weighs only about 4.5 billion parameters, so ready-made compressed versions exist for running it without a serious GPU. The license is Apache 2.0, and the dataset it was trained on is published separately under MIT, so both the model and what it learned from can be inspected and reused.
Why a vibe-coder should care
For anyone who wants a working 3D scene out of a description or a photo without opening a full editor, this is a compact model that genuinely runs on an ordinary laptop.
A regular laptop with 8 GB of RAM or more
Open on Hugging Face▌ More finds