Hugging Face
AI Platform
Open-source ML platform, model hub, Transformers library.
Watched repos
Recent activity
-
Hugging Face has partnered with OpenRouter to make open-weight models easy to run, allowing users to access the best provider for each request based on live measurements of price, speed, and reliability. This integration enables users to run open-weight models with the same API and token as before, while OpenRouter serves the models at scale across 80+ providers. The partnership aims to reduce costs and improve performance for open-weight models, which now cost about four times less on average compared to closed models. AI summary
Read more → -
To build a robot pick-and-place simulation in Isaac Sim, download the UR17G arm, Robotiq 2F-140 gripper, and beverage can assets from SimReady Central, which includes simulation settings and example scenes. Validate the assets by checking joint settings, movement, and contact behavior before running a pick-and-place task. Use an agent, such as Isaac Sim's remote Python server, to interact with the simulation and generate task code for the pick-and-place task. AI summary
Read more → -
XYBench is a benchmark of 8K+ questions that tests how large language models respond to requests based on misconceptions, measuring their ability to identify and address the underlying goal. The benchmark evaluates models' ability to recognize and redirect users toward effective solutions, focusing on pragmatic emphasis and misconception identification. The results show that current models often prioritize the stated request over the underlying goal, highlighting a need for improvement in model evaluation and training methods. AI summary
Read more → -
The llama.cpp v1.0.0 release provides a stronger foundation for local AI ecosystems, enabling multi-modal LLMs to run on a wide range of hardware. The library supports over 150 architecture variants and all mainstream speculative decoding methods, with a focus on quantization, optimized backends, and a portable C/C++ implementation. This release is part of a larger effort to make AI inference accessible at every level of the software stack. AI summary
Read more → -
Researchers have developed LLMON, a markup language for LLMs to explicitly communicate structure and semantics, addressing the brittleness and vulnerability to prompt injection attacks of current LLMs. The LLMON models, derived from Granite 4.1 base models, maintain general capability while eliminating prompt injection attacks by 47% and 95% compared to non-LLMON models on benchmarks. LLMON enables explicit instruction/data separation, indirect prompt injections are mostly ignored, and unrelated passages do not get pulled into the answer. AI summary
Read more → -
This article provides a fine-tuning guide for 3D generation models, specifically TRELLIS.2, using low-rank adaptation (LoRA). The model is fine-tuned on a subset of the Google Scanned Objects (GSO) dataset, which has 188 footwear objects. The fine-tuning process involves adding LoRA adapters to the sparse-structure flow DiT and the shape flow DiT, which control the object's shape and geometry. The adapters are trained using a rectified flow matching objective, and the original model weights are kept frozen. The fine-tuning script and containers are available for local and Hugging Face Jobs deployment. AI summary
Read more → -
BotPsy, a 0.8B-parameter agentic language model, achieves state-of-the-art (SOTA) tool-use and agentic performance for its size, outperforming models up to 1.2B parameters. It excels in instruction following, multi-turn conversations, and multi-step tasks, but struggles with open-ended search and complex information retrieval. BotPsy is designed for secure on-device workflows, providing exceptional instruction-following reliability and efficiency while keeping user data private. AI summary
Read more → -
Token pooling-aware training for multi-vector embedding models improves retrieval quality by teaching the encoder to preserve information when its vectors are merged, helping to retrieve better from the same storage budget. By applying a loss wrapper to a multi-vector retrieval loss, pooling-aware training can be used with existing models, resulting in a 97.8% retention of unpooled NanoBEIR nDCG@10 with a 4x reduction in document vectors. AI summary
Read more → -
Decision models in Hugging Face are a type of probabilistic model that evaluates a set of predefined questions about a given state and returns a probability distribution over possible decisions, without generating any text. They differ from classification models in that they can accept multiple types of questions (choice, score, noul) and return a structured decision along with probabilities. AI summary
Read more → -
You can now pay for Hugging Face Hub usage directly through your Google Cloud account, eliminating the need for a separate credit card. This integration allows Hugging Face usage to be billed alongside other Google Cloud services on a single invoice, streamlining billing and procurement processes. To set up this integration, follow these steps: 1) Subscribe to Hugging Face on the Google Cloud Marketplace, 2) Link your Hugging Face organization to your Google Cloud account, and 3) Have your subscription approved. AI summary
Read more → -
A synthetic dataset, Nemotron-Personas-Finance-Japan, has been released to capture Japan's demographics and financial context, comprising 9M personas with 50 fields each, including demographics, household finances, and financial literacy. The dataset is designed to support practical AI development in Japan, particularly in the financial sector, and is available for commercial use under the CC BY 4.0 license. It can be used to generate human-like dialogue, build domain-specific AI agents, and evaluate bias in AI models. AI summary
Read more → -
A new expert pruning method called HOPE (Higher-Order Pruning of Experts) is introduced, which improves the performance of mixture-of-experts (MoE) language models by accounting for pairwise expert interactions. HOPE achieves state-of-the-art results across pruning rates, with the highest rate of top-1 finishes (39%) and top-2 finishes (67%). The method is efficient, requiring only a calibration pass and solving a quadratic program, and can be applied to various MoE architectures. AI summary
Read more → -
Researchers have developed ToyBlocks, a controlled environment for robot learning and evaluation, enabling the creation of simulated worlds for studying robotics questions. The framework pairs a controlled, simulated world with tools for automatically generating tasks and training demonstrations, as well as metrics for measuring a robot's behavior over time. AI summary
Read more → -
Hugging Face has added Egyptian and Nigerian English to the Open ASR Leaderboard, a private dataset featuring scripted and conversational speech from five accents, including the US, UK, Australia, Canada, India, Nigeria, and Egypt, with richer accent breakdowns based on Kachru's Three Circles. AI summary
Read more → -
NVIDIA has released four new open-weight checkpoints for its Cosmos 3 model family, enabling the generation of interactive world simulations and multi-sensor multi-view driving scenes. These checkpoints, Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer, and Cosmos3-Nano-Sim-Bimanual, use OpenMDW1.1 and have 16B trainable parameters, allowing for commercial and non-commercial use. AI summary
Read more → -
Phonon, an on-device text-to-speech (TTS) model, has been released, allowing developers to run high-quality speech synthesis on their own devices without relying on servers or networks. The model achieves impressive results, ranking first in German and Spanish intelligibility and top three in all languages, while also outperforming larger models in speed and CPU streaming benchmark. AI summary
Read more → -
A 2-billion-parameter model, Dharma-AI/Dharma-OCR-LITE, extracts structured fields from any document type, outperforming general-purpose LLMs on field-level F1 score and Levenshtein similarity, while costing 6-147 times less than frontier APIs. The model achieves this by using a LoRA adapter on top of DharmaOCR-LITE v1.1, trained on 43,384 examples covering multiple document types and languages. The model's cost-effectiveness and quality are comparable to those of frontier APIs, but with a lower cost per extraction. AI summary
Read more → -
Researchers developed StateBench to test how language models track state changes over long documents, finding that recurrent and hybrid models can generalize much farther than transformers on simpler state-tracking problems, while transformers still struggle on harder forms of state tracking. StateBench consists of four difficulty levels, capturing increasingly complex state update operations. The benchmark showed that models trained directly on these tasks can handle state-tracking levels their pretrained counterparts couldn't, and highlights the need for a finer-grained understanding of state-tracking capabilities in language models. AI summary
Read more → -
A new model, Palette, was trained to build decks in a design language, competing with larger models like Frontier, and achieved comparable results. Palette's approach involves a small, open-weight model trained on synthetic data and checked by a harness, which allows for finer-grained control over design choices and enables the model to learn from its mistakes. AI summary
Read more → -
The article introduces Model Integration Contracts (MICs) as a way to test and validate Hugging Face models, allowing for the creation of smaller, more efficient models that can catch bugs before they reach the mainline codebase. MICs are built using a "shrink" recipe that reduces the model's configuration while keeping its architecture intact, and are tested using a set of predefined usage paths and baselines. AI summary
Read more →