Top 5 Hugging Face Models This Week

I put Clef from Cloudflare on my own test support tickets and gave it several questions at once: is the client angry, does it need escalation, what type of problem is it. The model does not write an answer, it returns a probability for every option right away, and it does this in one pass with 27 billion parameters. By Cloudflare's own table, on the BANKING77 test (sorting of banking requests) Clef scores 94.2 against 79.7 for the paid Jev, which it competes with directly. I liked that it does not get confused on narrow sorting, but for questions that need general knowledge I would not choose it, there the paid Jev wins by a wide margin.

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

I downloaded LTX-2.5 from Lightricks and ran a couple of prompts through it for a video with several scenes in a row. The model keeps the characters, the light, and the voices the same across scenes, and this is the main difference from the previous version. Instead of the usual VAE there is a diffusion decoder here, the picture is noticeably cleaner, and on-screen text is readable. Ready workflows ship right inside ComfyUI, so I did not have to assemble everything piece by piece.

💻 A GPU with 24 GB of VRAM or a Mac with 32 GB of memory; for weaker cards there is a lighter offload mode.

I set up Qwen3.8-27B from Alibaba locally and gave it a screenshot of code plus a question in Russian. It natively understands images and video, not only text, and it can decide on its own whether to think fast or think long before answering. Compared with the previous Qwen3.6-27B the gain is noticeable, on the SWE-bench Pro coding test it is 61.7% against 53.5%. For local use there are already more than 700 quantized versions, including an official one from unsloth, so setting it up is not hard.

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

I downloaded d1-3B from LiquidAI, only 3.1 billion parameters, and tried it on comment moderation: spam, toxic, on topic. It also does not write text, it returns a calibrated probability for every question right away, including on images, not only on text. By LiquidAI's own test (Decision Index) it beats models with 4 and even 9 billion parameters. On an RTX 4090 the answer takes 8 milliseconds, on an Apple M5 Pro chip 30, and I liked that it does not need a powerful graphics card for this speed.

💻 A regular laptop with 8 GB of RAM or more.

I set up Laya from Convai Innovations (the handle on Hugging Face is convaiinnovations) and ran a few test support emails through it with standard questions. It also answers not with text but with a probability in one step, the English version has 421 million parameters, and a separate multilingual version with 322 million covers over 100 languages. On a GPU the answer to one question takes 32.8 milliseconds, the paid Jev needs 236 to 276 milliseconds for the same question, and on a plain CPU, without a graphics card, Laya needs 0.2 to 0.5 seconds. The one thing that bothers me is that on questions with 20 or more answer options it starts to confuse them, and here fine-tuning on your own data raises accuracy from 0.362 to 0.766.

💻 A regular laptop with 8 GB of RAM or more, no graphics card needed.