From Hugging Face Blog · 28 stories
Direct Preference Optimization Reduces Text Degeneration in OCR Models
DharmaOCR introduces Direct Preference Optimization (DPO) to combat text degeneration in OCR models. The second training stage reduced degeneration rates by an average of 59.4%, addressing a significant limitation of supervised fine-tuning.
Holo3.1 Released with Local Execution and Enhanced Performance
Holo3.1 has been released, featuring enhanced robustness for local and mobile environments, quantized checkpoints for local inference, and improved performance across various deployment frameworks. This release addresses the challenges of deployment flexibility and performance consistency in diverse operational settings.
JetBrains Launches Mellum2: 12B Mixture-of-Experts AI Model
JetBrains has released Mellum2, a 12 billion-parameter Mixture-of-Experts model optimized for natural language and coding tasks. With efficient parameter activation and over 2x faster inference compared to similar models, Mellum2 is positioned for high-throughput AI applications.
New Research Addresses LLM Safety Refusal for Specific Harmful Subsets of Topics
New research explores how large language models (LLMs) can be trained to refuse only specific harmful subsets of a topic, rather than the entire topic. This approach allows LLMs to maintain helpfulness while still adhering to safety policies, addressing a limitation where current topic-level safety guards are too broad.
Constraint-Aware GPU Allocator Improves Utilization by 33 Points Over FIFO Scheduling
A new constraint-aware GPU allocator demonstrated up to a 33 percentage point increase in GPU utilization and up to 105% increase in priority-weighted output compared to a FIFO scheduler on identical hardware and workloads. This improvement highlights the significant impact of allocation order on resource efficiency in GPU clusters, particularly when managing diverse workload types like training and real-time inference.
Open Model Landscape in Summer 2026 Shows Rapid Growth and Shifting Development Strategies
The open model ecosystem on Hugging Face grew significantly in 2026, with models increasing from 2.43 to 2.96 million and datasets from 711,000 to 1 million. Chinese labs are increasingly releasing large, high-performance models, often exceeding those from American labs, while hardware vendors like AMD and NVIDIA are becoming major contributors of open models to drive chip sales.
Challenges and Advances in Simulation for Physical AI Systems
The article discusses the challenges of data availability in training physical AI systems and highlights the role of simulation in overcoming these issues. Simulation enables the generation of photorealistic data at lower costs, allowing developers to enhance robot learning and performance in complex physical interactions.
Routing Systems in AI: Complexity Beyond Model Selection
Routing systems for AI agents face complexity beyond simple model selection, involving cost, performance, and compliance challenges. Caching effects and task difficulty assessments must also be factored into routing decisions for optimal efficiency.
Analysis of AI Specialization and Its Emergence as a Key Principle
A recent analysis highlights the inevitability of specialization in effective AI systems, drawing on various domains. It argues that focused AI systems outperform general models, correlating with findings in optimization theory and evolutionary biology.
Exploring Alternatives to LoRA in Parameter-Efficient Fine-Tuning
The article investigates alternatives to LoRA, the predominant technique in parameter-efficient fine-tuning (PEFT). It highlights the potential of PEFT techniques to reduce memory requirements for model fine-tuning and mentions the development of the PEFT library by Hugging Face, which supports various methods and improves accessibility.
Hugging Face Models Now One-Click Deployable to Amazon SageMaker Studio
AWS and Hugging Face have integrated deep-linking, allowing developers to move a model from Hugging Face directly into Amazon SageMaker Studio in a single click. This eliminates prior multi-step processes, enabling quicker model experimentation and deployment. The update is significant for faster AI development and deployment in enterprise environments.
Open ASR Leaderboard Adds Egyptian and Nigerian English Accents
The Open ASR Leaderboard has expanded its private data sets to include Egyptian and Nigerian English accents through a collaboration with Liva AI. This addition aims to improve representation in Automatic Speech Recognition (ASR) evaluations and provide more diverse data for model testing.
NVIDIA Nemotron Voice Agent Powers Interactive Halloween Projection
An interactive Halloween projection named MonsterHouse was created using NVIDIA's Nemotron Voice Agent, Nemotron ASR, Nemotron 3.5 Bolt E4B 4GB, Magpie TTS Zeroshot, and Audio2Face-3D. The system allows a garage door to listen to trick-or-treaters, respond as a grumpy pirate, and animate a mouth. This project demonstrates the practical application of NVIDIA's AI tools for real-time conversational agents and animation in a creative context.
LitPivot System Aids Researchers in Developing Ideas by Contextualizing Literature
Researchers developed LitPivot, a system that helps researchers refine early-stage research ideas by organizing literature, offering critiques, and suggesting revisions. A study showed that ideas developed with LitPivot were better grounded in literature and perceived as more novel and useful compared to those developed with text editors and AI chat tools.
Falcon-Emirati-7B LLM Released for Emirati Arabic Dialect Understanding
A new large language model, Falcon-Emirati-7B, has been released, specifically designed to understand and generate Emirati Arabic, including its cultural nuances. This model builds upon the Falcon-H1-Arabic family to address the gap in LLMs for specific Arabic dialects beyond Modern Standard Arabic.
OlmoEarth Studio now offers custom embedding exports for Earth observation data
OlmoEarth Studio has introduced the ability to compute and export embedding vectors from its open-source OlmoEarth foundation models. These embeddings provide compact numerical representations of Earth-observation data, enabling various downstream analytical tasks such as similarity search and segmentation.
Hugging Face Expands PyTorch Profiling Guide with MLP and Attention Techniques
Hugging Face continues its 'Profiling in PyTorch' series, detailing the integration and profiling of nn.Linear and Multilayer Perceptron (MLP) blocks, and expanding to attention mechanisms in transformer models. These insights assist developers in optimizing deep learning models using the PyTorch profiler, showcasing GPU capabilities effectively.
DiScoFormer model estimates density and score for data distributions
The DiScoFormer model estimates both the density and score of data distributions in a single forward pass. This model improves upon existing methods by allowing for high-dimensional data analysis without the need for retraining, addressing challenges in density estimation and score matching.
Hugging Face simplifies vLLM server setup with single command
Hugging Face introduced a command to run a vLLM server easily, facilitating model testing and evaluation. This command allows users to quickly deploy models and interact with them via the OpenAI API using Hugging Face infrastructure.
Hugging Face Enhances CLI and Adopts Weekly Releases for Improved Efficiency
Hugging Face has updated their command-line interface (CLI) to cater to both human and artificial intelligence (AI) agents, optimizing token usage. Additionally, they have shifted to a weekly release schedule for the huggingface_hub Python client to accelerate the implementation of fixes and features. These changes enhance CLI efficiency and streamline the release process.
Strands Robots SDK integrates LeRobot for seamless robot task management
The Strands Robots SDK now integrates LeRobot hardware and simulations, streamlining task management for robots. Users can record, test, and deploy robot tasks with fewer tools, enhancing workflow efficiency across multiple robots.
Agent Creates 3D Paris Gallery Using Hugging Face Spaces
A coding agent utilized Hugging Face Spaces to create a web gallery featuring 3D Gaussian models of Paris monuments without manually engaging with image or 3D tools. This illustrates a shift towards modular software construction where AI integrates existing components easily.
Introduction of MCP Tools for Reachy Mini Enhances Remote Functionality
The Reachy Mini now supports remote tools through MCP canary Space, allowing the addition of external functionalities like weather queries. This update enhances the robot's interactivity and potential use cases without modifying the core app directly.
GPU Utilization Becomes Key Constraint for Enterprise AI, Similar to Airline Aircraft Downtime
The efficiency of GPU utilization is emerging as a critical factor for enterprise AI success, mirroring how aircraft ground time impacts airline profitability. As AI scales, the focus shifts from model quality and raw compute power to maximizing the active use of specialized hardware to control costs and drive output.
Build a Robot Pick-and-Place Simulation with SimReady Assets in Isaac Sim
A walkthrough demonstrates building a robot pick-and-place simulation in Isaac Sim using SimReady assets. This process involves composing a scene with a UR17G arm, Robotiq 2F-140 gripper, and a beverage can, then checking their behavior and running a pick-and-place example with agent-generated controls. SimReady assets are OpenUSD assets prepared with properties like collisions, mass, and joints for specific simulation use cases.
Fine-Tuning Guide and Scripts Released for TRELLIS.2 3D Generation Model
Fine-tuning scripts and containers are now available for TRELLIS.2, a 3D generation model, allowing users to adapt it to specific 3D assets. These resources can run locally or on Hugging Face Jobs, addressing the need for task-specific accuracy in 3D model generation.
Open-source reproduction of AI model for generating watercolor paintings with TRL and OpenEnv
A developer has openly reproduced the training pipeline for an AI model that generates watercolor paintings using JavaScript and the p5.brush library. This reproduction utilizes Hugging Face's TRL and OpenEnv, making the training scripts, RL environment, and models publicly available.
Finetuning Multi-Vector Embedding Models with Sentence Transformers
A guide details how to finetune multi-vector embedding models using Sentence Transformers, including a new `MultiVectorEncoder` for ColBERT-style late interaction retrieval. This method allows users to train models that can outperform general-purpose retrievers on specific datasets.