From Hugging Face Blog · 40 stories
Holo4 Agentic Models Released, Featuring Generalist Computer-Use Capabilities
H Company released Holo4, a new series of agentic models available in 27B dense and 35B-A3B Mixture of Experts sizes, along with Holotron4 Nano. These models interact with software across GUIs, code, MCP, and APIs, aiming to handle diverse business workflows by combining different interface approaches.
NVIDIA Releases Nemotron 3 Diarization Model for Multi-Speaker AI
NVIDIA has released Nemotron 3 Diarization, an open-weight, 100M-parameter model designed to identify who spoke when in multi-speaker conversations. This model improves upon previous versions by supporting up to eight speakers and achieving a 14.72% Diarization Error Rate (DER) on VoiceArena's Diarization-Bench leaderboard, making it relevant for applications requiring accurate speaker attribution in complex audio.
AISI and EvalEval Release Reproducible AI Benchmark Results on Evaluation Cards
The UK's AISI and EvalEval have released publicly reported evaluation methods and findings through EvalEval's Evaluation Cards platform, including verified results for five benchmarks and six frontier models. This collaboration aims to improve the reproducibility and transparency of AI model evaluations by providing a shared infrastructure for reporting and interpreting results.
oMLX Creator Jun Kim Joins Hugging Face to Support MLX Community Development
Jun Kim, the creator and maintainer of oMLX, has joined Hugging Face to continue leading and developing the oMLX project. This move aims to provide more stability and faster development for oMLX, an open-source project focused on local AI optimized for Apple Silicon.
Multiverse AI uses Ising model for LLM pruning, improving Llama-3.3-70B-Instruct compression
Multiverse AI developed a new method for pruning large language models (LLMs) by reformulating block removal as a constrained binary optimization problem, mapping it to an Ising glass model. This approach allows for ranking candidate configurations without extensive benchmarking, achieving a 23 percentage point gain on MMLU for Llama-3.3-70B-Instruct at 50% compression compared to other block-removal methods. The method accounts for inter-block dependencies, which traditional mean-field approaches overlook.
Hugging Face TRL v1.14 AsyncGRPOTrainer Adds LoRA Support for Distributed RL Training
TRL v1.14's AsyncGRPOTrainer now supports LoRA adapter training, allowing only the small adapter to be synced to vLLM. This enables distributed reinforcement learning setups where the trainer and vLLM replicas run as separate Hugging Face Jobs on different machines, improving training efficiency.
Gradio Workflow Rebuilds AUTOMATIC1111 Features as a Single Workflow Canvas
Gradio has rebuilt most of AUTOMATIC1111's stable-diffusion-webui features into a single workflow canvas called Workflow1111. This allows users to run eleven media pipelines, including text-to-image, hi-resolution fix, and image-to-image, using a graph of 73 nodes within the Gradio environment.
IBM releases Granite Time Series PatchTST-FM-r2 model with commercial-friendly license
IBM has released Granite Time Series PatchTST-FM-r2, an updated time-series foundation model with approximately 385 million parameters, under Apache 2.0 and OpenMDW 1.0 licenses. This model achieves top zero-shot performance among permissively licensed, replicable models on the GIFT-Eval benchmark, offering capabilities like probabilistic forecasting and missing value imputation.
IBM Time Series Foundation Models Now Available on Confluent Cloud for Real-Time Intelligence
IBM is integrating its time series foundation models (TSFMs) with Confluent Cloud, allowing businesses to apply advanced forecasting, anomaly detection, and optimization directly to real-time data streams. This collaboration enables domain experts to utilize AI models for operational decisions without requiring extensive data science expertise, potentially improving accuracy and productivity in various industries.
BenchMIRT Introduced to Audit LLM Benchmarks at the Prompt Level
BenchMIRT is a new method for auditing large language model (LLM) benchmarks by analyzing individual prompts to determine which underlying capabilities they measure. This tool helps researchers understand what drives a benchmark's score by separating multiple contributing abilities, rather than relying on a single averaged score. It matters because current LLM benchmarks can obscure the specific abilities being tested, making it difficult to accurately assess model performance.
Hugging Face Releases @huggingface/kernels for WebGPU-Optimized Local AI
Hugging Face launched @huggingface/kernels, a library providing over 200 optimized WebGPU kernels for local AI operations, alongside Fleet, an in-browser GPU benchmarking suite. This initiative aims to standardize and optimize WebGPU-based machine learning computations across diverse hardware, improving performance and portability for AI models running in browsers.
Open ASR Leaderboard Adds Hindi and Indian English to Address Bias in Speech Recognition
The Open ASR Leaderboard has integrated two new evaluation sets, Monsoon en-IN and Monsoon hi-IN, to include Hindi and Indian English, marking the first Global South languages on the platform. This addition aims to address known biases in automatic speech recognition (ASR) systems, which often perform poorly for non-European languages and diverse speaker demographics.
Gradio Introduces gr.Workflow for Building AI Pipelines with Visual Interfaces and REST APIs
Gradio has launched gr.Workflow, a new feature that allows users to define AI pipelines as a graph of typed nodes, providing a drag-and-drop canvas for interaction. This development simplifies the creation and deployment of multi-step AI applications, making them accessible as both visual interfaces and REST APIs, and enabling one-command deployment to Hugging Face Spaces.
Hugging Face Storage Buckets Integrate with Strands Agents and LeRobot for Continuous AI Training
Hugging Face Storage Buckets, a new mutable object-storage repository type, now integrates with Strands Agents and LeRobot to facilitate continuous recording, training, and deployment of robot policies. This integration addresses data transfer inefficiencies in iterative AI development by providing a working layer for data between recording and training phases.
Baseten Integrates as an Inference Provider on Hugging Face Hub
Baseten is now a supported Inference Provider on the Hugging Face Hub, allowing developers to use Baseten's serverless AI platform for conversational and text-generation tasks directly from Hugging Face model pages and SDKs. This integration expands the options for deploying and utilizing open-weight large language models within the Hugging Face ecosystem.
AI2 Launches OlmoEarth Platform for Geospatial AI Inference at Planetary Scale
AI2 has launched the OlmoEarth Platform, an infrastructure designed to facilitate large-scale geospatial model inference using its OlmoEarth foundation models. This platform addresses the challenges organizations face in deploying Earth observation AI, enabling applications like deforestation monitoring and wildfire risk assessment.
Hugging Face Diffusers Now Natively Supports Nunchaku 4-bit Diffusion Inference
Hugging Face Diffusers now natively supports Nunchaku 4-bit diffusion inference, which utilizes SVDQuant for 4-bit weights and activations. This integration allows users to run diffusion models with reduced memory usage and faster inference directly within Diffusers, eliminating the need for separate inference libraries or local CUDA compilation.
Grabette Launches Open System for Recording Robot-Manipulation Data
Grabette is a new open-source system designed to simplify the recording of robot-manipulation data. By allowing users to capture demonstrations using a handheld device without requiring a robot, it aims to democratize data collection and support the development of robust robot learning models.
DharmaOCR Shows Superior Performance in Brazilian Portuguese OCR
DharmaOCR outperformed Mistral OCR4 and Unlimited-OCR in Brazilian Portuguese OCR through targeted training. The model's training process enhanced extraction quality and reliability, addressing known challenges in the technology.
Skylight's Shippy AI Enhances Maritime Domain Awareness Reliability
Skylight has developed Shippy, an AI agent designed for real-time maritime domain awareness, prioritizing reliability. Shippy integrates live data and provides verified responses to maritime analysts, helping to minimize the risks of incorrect information in high-stakes operations.
Real World VoiceEQ Benchmark Launches To Evaluate Voice AI Quality
Real World VoiceEQ has been introduced as a benchmark for assessing the human quality of voice models. The tool measures 15+ evaluation dimensions and identifies failures in voice interactions, aiming to improve the reliability of voice AI systems in real conversations.
vLLM Enhances Transformers Integration for Optimized Model Inference
The vLLM pip package now features improved integration with the transformers library, allowing users to run Hugging Face models more efficiently. This update introduces advanced inference techniques that optimize performance across various model sizes and architectures.
SkyPilot integrates with Hugging Face for zero-egress AI workload storage
SkyPilot now supports direct integration with Hugging Face, allowing users to run AI workloads on any cloud without egress fees. This integration enables seamless access to models and datasets stored on Hugging Face, optimizing cloud compute resources across various providers.
PRX Update: New Data Strategy for Training Model
PRX outlines its data strategy for model training, focusing on assembling a diverse dataset. The approach emphasizes breadth over perfection in data selection, utilizing existing public and internal datasets for training efficiencies.
LeRobot v0.6.0 Launches with New World Models and Reward APIs
LeRobot v0.6.0 introduces world model policies and new reward models APIs, alongside updates to datasets and benchmarks. These enhancements aim to improve robotic training and evaluation, enabling more efficient simulations and real-world applications.
Hugging Face Introduces New Kernel Repository with Enhanced Security Features
Hugging Face has launched a new repository type called 'kernel' to improve discoverability and security for compute-oriented users. The platform now includes layers of protection against malicious code and features such as trusted publishers and code signing.
ScarfBench Launches as New AI Benchmark for Java Framework Migration
ScarfBench provides a new open benchmark to evaluate AI agents on Enterprise Java framework migrations. It focuses on ensuring successful builds, deployments, and behavior preservation across major Java ecosystems like Spring and Jakarta EE, addressing gaps in existing AI-assisted modernization efforts.
Hugging Face Integrates Every Eval Ever for Model Reporting
Hugging Face has integrated the Every Eval Ever (EEE) JSON schema into its Community Evals to standardize AI evaluation reporting. This collaboration aims to enhance trust and comparability in model performance, addressing inconsistencies in evaluation results reported across multiple formats.
Hybrid models outperform transformers in predicting meaning-rich tokens
Experiments revealed that hybrid models, like Olmo Hybrid, predict meaning-rich tokens better than transformers. However, on simple repetitive tokens, transformers maintain an edge, indicating differing strengths in architectural approaches.
NVIDIA NeMo AutoModel Enhances Fine-Tuning for Generative AI Models
NVIDIA launched NeMo AutoModel, enhancing fine-tuning for generative AI models by enabling higher training performance. This tool achieves up to 3.7x faster training and reduces GPU memory use by up to 32%, making it easier for developers to implement advanced models without extensive code changes.
Launch of FFASR Leaderboard to Benchmark ASR in Real-World Conditions
Treble Technologies and Hugging Face introduced the FFASR Leaderboard to evaluate Automatic Speech Recognition (ASR) models under far-field conditions. This community-driven benchmark aims to address the significant gap in performance between traditional clean-speech evaluations and real-world usage scenarios involving background noise and reverberation.
IBM's CUGA Offers Lightweight Framework for Building Agentic Apps
IBM has released CUGA, a Configurable Generalist Agent harness that simplifies the development of agentic applications by automating the orchestration and state management. With two dozen example applications provided, developers can create functional agents quickly without extensive groundwork, increasing efficiency in building machine learning applications.
Transformers.js improves browser-based AI model management with Cross-Origin Storage API
Transformers.js now integrates the proposed Cross-Origin Storage API to manage AI model resources more efficiently. This change reduces the redundant downloads of commonly used models across different web applications, addressing issues related to cache storage and data usage.
PP-OCRv6: New OCR Model on Hugging Face with 50-Language Support
PaddleOCR has launched PP-OCRv6, a new OCR model with capabilities in 50 languages and scalability from 1.5M to 34.5M parameters. The model improves text detection and recognition accuracy compared to its predecessor, PP-OCRv5, making it suitable for a variety of real-world OCR applications.
Local Models Triaged Issues in OpenClaw Repository for Free
In June 2026, local AI models were used to efficiently triage issues in the OpenClaw repository. This method allows for real-time notifications and reduces costs associated with cloud-based models, highlighting the growing importance of local AI implementation.
MosaicLeaks addresses privacy risks in deep research agents with new training method
MosaicLeaks reveals privacy vulnerabilities in deep research agents that combine private documents and web searches, leading to potential leakage of sensitive information. The proposed Privacy-Aware Deep Research (PA-DR) method improves task accuracy and decreases information leakage significantly, from 34.0% to 9.9% for full-information leakage.
New ARD Specification Enables Dynamic Agent Searches Across Tools
The Agentic Resource Discovery (ARD) specification has been developed collaboratively by major tech companies to allow agents to discover tools at runtime. This move shifts from a static model requiring pre-installed capabilities to dynamic, intent-based searches, enhancing the ability of agents to access and utilize a broader range of tools effectively.
Migrating CI from GitHub to Hugging Face Jobs for Enhanced Performance
Trackio has migrated its CI from GitHub Actions to Hugging Face Jobs, achieving a 30% reduction in CPU CI time and enabling GPU testing. This step is significant for improving efficiency and expanding testing capabilities in machine learning projects.
OpenEnv Gains Support from Major AI Organizations for Open Source Development
OpenEnv has transitioned to an open-source model coordinated by leading AI organizations such as Meta-PyTorch and Microsoft. This move aims to improve agent training efficiency across various AI harnesses and environments, fostering collaboration within the AI community.
Nemotron 3.5 Enhances Multimodal Content Safety with Custom Policies
Nemotron 3.5 introduces customizable multimodal safety integration, considering user prompts, images, and responses simultaneously. This update captures policy violations emerging from interaction, enhancing deployments across various global languages and industries.