The top 20 hugging face models measured by RiseFinder over the 30d window ending September 2026.
Turns ordinary homes into calibrated recording studios, capturing spatially aligned and time synchronised data for training robots in real domestic settings rather than in a lab.
Hugging Face models
Open source ↗A LoRA for the MiniMax H3 video model tuned for realistic human figures.
Hugging Face models
Open source ↗A curated collection of 15,458 open and partially solved mathematics problems spanning every domain and difficulty level, including what its authors describe as the largest assembled set of Erdős problems. It exists as training and evaluation material for mathematical reasoning models, where the scarce resource is not proofs but problems that are genuinely unsolved, since anything already answered risks sitting in the training data.
Hugging Face models
Open source ↗A vision language model built on the Qwen3.5 mixture of experts architecture, able to hold an image grounded conversation and take text and images as input.
Hugging Face models
Open source ↗1,080,814 images cut from 49,455 digitised books held by the British Library.
Hugging Face models
Open source ↗A deduplicated training corpus of traces from three frontier models across maths, code and reasoning, filtered for quality.
Hugging Face models
Open source ↗A small evaluation dataset created to probe what the MiniMax H3 model knows and what kinds of tasks it can handle. It is intended for capability testing rather than as a broad training corpus or a benchmark tied to one application.
Hugging Face models
Open source ↗A collection of spoken character lines from Genshin Impact assembled as an audio dataset. It provides organised voice material for experiments involving speech recognition, speaker modelling and other audio research workflows.
Hugging Face models
Open source ↗A GGUF conversion of Liquid AI's 2.6 billion parameter model, for running on a device through llama.cpp.
Hugging Face models
Open source ↗A large mathematical reasoning dataset from NVIDIA containing model generated solution traces, for fine tuning smaller models on how to work a problem through.
Hugging Face models
Open source ↗A dataset of first person video capturing real hand movements and object manipulation, paired with 3D hand pose reconstructions, camera position tracking and labelled actions for each moment. Around 1,200 of a planned 2,000 hours have been released so far, aimed at training and evaluating robots or models that need to understand how a human hand actually manipulates objects.
Hugging Face models
Open source ↗Raw unscripted first person video of daily life.
Hugging Face models
Open source ↗An 11 billion parameter speech model from NVIDIA that handles listening, understanding and speaking in one unified system rather than chaining separate speech recognition, language model and text to speech components. It responds in roughly 450 milliseconds, can be interrupted mid response, and can call external tools while still holding a natural conversation.
Hugging Face models
Open source ↗The structured output and function calling data behind the Hermes 2 Pro models, released so others can train on the same material.
Hugging Face models
Open source ↗FLUX.1 dev, Black Forest Labs' open-weight text-to-image model. The non-commercial research release of the family that became the default open image model after Stable Diffusion.
Hugging Face models
Open source ↗A mixture of experts model using ternary weights, which cuts memory to a fraction of a normal float model.
Hugging Face models
Open source ↗Two thousand hours of robot manipulation footage from six synchronised camera views across 480 scenes, with pose accurate to three millimetres.
Hugging Face models
Open source ↗Close to a million synthetic personas, each described by 1,290 categorical attributes, for populating simulations and tests.
Hugging Face models
Open source ↗A 2.6 billion parameter model from Liquid AI, built on their own architecture and sized for running on a device rather than a server.
Hugging Face models
Open source ↗A speech dataset of more than 611,000 recordings, over 3,000 hours in total, covering thirteen Arabic dialects from countries including Egypt, Iraq, Lebanon, Morocco and Saudi Arabia. Transcriptions preserve each dialect's own spelling rather than normalising to Modern Standard Arabic, aimed at training speech recognition and dialect identification systems.
Hugging Face models
Open source ↗Get early signals on the websites, apps, repositories and startups gaining momentum. Join the list for Risefinder email briefings.
Be first to receive the briefing.