Articles from Source: Hugging-Face-Blog

Build an Agent That Thinks Like a Data Scientist: How We Hit #1 on DABStep with Reusable Tool Generation

2026-03-13 01:02
Unlocking the potential of data analysis, the NVIDIA KGMON team has developed the NeMo Agent Toolkit Data Explorer. This innovative agent is designed for dataset exploration, tackling complex tasks like multi-step reasoning and iterative analysis. Their approach achieved #1 ranking on the DABStep benchmark, showcasing a remarkable 30x speedup compared to existing methods. 🚀📊 #DataScience #AI #NVIDIA #KGMON #Innovation
Source: Hugging Face Blog

How NVIDIA AI-Q Reached \#1 on DeepResearch Bench I and II

2026-03-12 03:53
NVIDIA's AI-Q has achieved top rankings on both DeepResearch Bench I and II, scoring 55.95 and 54.50, respectively. 🥇 This success highlights the potential of open, portable deep research tools. AI-Q offers a modular architecture that allows enterprises to customize AI agents for their specific needs. Learn more about how AI-Q is shaping the future of AI-driven research! 🔍✨ #NVIDIA #AI #DeepResearch #Innovation #TechNews
Source: Hugging Face Blog

Code Concepts: A Large-Scale Synthetic Dataset Generated from Programming Concept Seeds

2026-03-11 15:50
Introducing "Code Concepts," a new synthetic dataset aimed at enhancing programming skills in large language models (LLMs). This dataset, created from programming concept seeds, contains 15 million Python problems. It addresses the need for high-quality, concept-targeted data in LLM development. The inclusion of this dataset in the Nemotron-Nano-v3 pretraining led to a notable improvement, achieving a six-point gain on the HumanEval benchmark. 📈💻 #DataScience #MachineLearning #Programming #AI...
Source: Hugging Face Blog

How NVIDIA Builds Open Data for AI

2026-03-10 19:49
NVIDIA emphasizes the importance of open data in developing trustworthy AI systems. 🤖 As AI models evolve, the quality of training data becomes crucial for their performance and safety. However, much of this data is still fragmented and hard to access. NVIDIA addresses this by providing open datasets, allowing developers to build high-quality models more efficiently. 📊 #AI #OpenData #NVIDIA #MachineLearning #DataAccess
Source: Hugging Face Blog

Introducing Storage Buckets on the Hugging Face Hub

2026-03-10 00:00
🚀 Exciting news from Hugging Face! They have introduced Storage Buckets on the Hugging Face Hub, designed specifically for production ML needs. These buckets allow for mutable, S3-like object storage, perfect for handling intermediate files like checkpoints and logs. You can easily browse, script from Python, or manage these files using the hf CLI, streamlining your workflow. #HuggingFace #MachineLearning #StorageBuckets #AI #DataManagement
Source: Hugging Face Blog

Keep the Tokens Flowing: Lessons from 16 Open-Source RL Libraries

2026-03-10 00:00
Exploring the world of open-source reinforcement learning (RL) libraries reveals valuable lessons. The article examines 16 RL libraries, highlighting their unique features and contributions to the field. It emphasizes the importance of community engagement and collaboration in advancing technology. These insights can guide developers and researchers in optimizing their own projects. 🔍💻📚 #ReinforcementLearning #OpenSource #TechInsights #AI #MachineLearning
Source: Hugging Face Blog

Granite 4.0 1B Speech: Compact, Multilingual, and Built for the Edge

2026-03-09 18:36
🚀 Introducing Granite 4.0 1B Speech by IBM! This new speech-language model is designed for enterprise use on resource-limited devices. It features multilingual automatic speech recognition (ASR) and bidirectional speech translation (AST) with improved performance. With reduced parameters, it offers enhanced English transcription accuracy and faster processing. Notably, it adds Japanese ASR support and keyword list biasing for better recognition. Granite 4.0 1B Speech has also topped the...
Source: Hugging Face Blog

LeRobot v0.5.0: Scaling Every Dimension

2026-03-09 00:00
🚀 Exciting news from the tech world! LeRobot v0.5.0 has just launched, showcasing over 200 merged PRs and contributions from 50+ new developers. This release features expanded capabilities, including support for the Unitree G1 humanoid robot and new training policies. Key highlights include faster datasets, simulation environments, and a modernized codebase using Python 3.12 and Transformers v5. #LeRobot #TechUpdate #Robotics #Innovation #AI
Source: Hugging Face Blog

Ulysses Sequence Parallelism: Training with Million-Token Contexts

2026-03-09 00:00
🚀 Training large language models is evolving! The article discusses Ulysses Sequence Parallelism, a method designed to handle long sequences of up to millions of tokens. This approach is crucial for tasks like document analysis and complex reasoning. Ulysses tackles memory challenges by distributing attention computation across multiple GPUs, making it easier to manage large contexts. It's integrated within the Hugging Face ecosystem, enhancing tools like Accelerate and the Transformers...
Source: Hugging Face Blog

Conversational LLM Evaluations in Minutes with NVIDIA NeMo Evaluator Agent Skills

2026-03-06 18:56
🚀 Exciting news for developers! NVIDIA has introduced the nel-assistant agent skill, simplifying LLM evaluations. Now, you can configure, run, and monitor evaluations using natural language, eliminating the need for complex YAML files. This innovation is built on the NVIDIA NeMo Evaluator library and is compatible with tools like Cursor. #NVIDIA #LLM #DeveloperTools #AI #NeMo
Source: Hugging Face Blog

Bringing Robotics AI to Embedded Platforms: Dataset Recording, VLA Fine‑Tuning, and On‑Device Optimizations

2026-03-05 14:16
🚀 Advances in AI are reshaping robotics! The article discusses recent developments in Vision–Language–Action (VLA) models, which integrate visual perception and robot actions. However, deploying these models on embedded platforms poses challenges due to limited compute and power resources. Additionally, the article highlights issues with synchronous control pipelines that can lead to oscillatory behavior and delayed responses. #Robotics #AI #EmbeddedSystems #Technology #Innovation
Source: Hugging Face Blog

Introducing Modular Diffusers - Composable Building Blocks for Diffusion Pipelines

2026-03-05 00:00
📢 Exciting news in the world of diffusion pipelines! The article introduces Modular Diffusers, a system that allows users to build diffusion pipelines using reusable blocks. This innovation enables the creation of customized workflows without starting from scratch. It also details the integration with Mellon, a visual interface for connecting these blocks. Key topics include quickstart guides, custom block creation, and community pipelines. 🔗 #ModularDiffusers #DiffusionPipelines...
Source: Hugging Face Blog

PRX Part 3 — Training a Text-to-Image Model in 24h!

2026-03-03 16:50
🚀 In the latest article, the process of training a text-to-image model in just 24 hours is explored. The team details the techniques and tools used to achieve this rapid training. 🔍 Key steps include data preparation, model selection, and optimization strategies that enhance performance. 💡 The article emphasizes the importance of efficient workflows in AI development. #AI #MachineLearning #TextToImage #Innovation #TechTrends
Source: Hugging Face Blog

Mixture of Experts (MoEs) in Transformers

2026-02-26 00:00
Discover the concept of Mixture of Experts (MoEs) in Transformers. This approach enhances model efficiency by activating only a subset of experts during inference, reducing computational resources. Researchers highlight the potential benefits of MoEs in improving performance without increasing the overall model size. Explore how this innovation could shape the future of AI development! 🤖📈 #MachineLearning #ArtificialIntelligence #Transformers #MoEs #AIResearch
Source: Hugging Face Blog

Deploying Open Source Vision Language Models (VLM) on Jetson

2026-02-24 00:00
🌟 Vision-Language Models (VLMs) combine visual perception with semantic reasoning, enhancing AI capabilities beyond traditional models. The NVIDIA Jetson family is designed to support these advanced models, offering optimized performance for edge devices. This tutorial outlines deploying the NVIDIA Cosmos Reasoning 2B model on Jetson using the vLLM framework. #AI #NVIDIA #Jetson #VLM #OpenSource
Source: Hugging Face Blog

GGML and llama.cpp join HF to ensure the long-term progress of Local AI

2026-02-20 00:00
🚀 Exciting news! GGML, the team behind Llama.cpp, is joining HF to promote the future of Local AI. Their collaboration aims to scale support for the community as Local AI evolves. With key contributors already involved, this partnership is set to enhance local inference and model definition. Stay tuned for more developments! 🤖 #LocalAI #Collaboration #AIFuture #OpenSource #Innovation
Source: Hugging Face Blog

Train AI models with Unsloth and Hugging Face Jobs for FREE

2026-02-20 00:00
Unlock the potential of AI training with Unsloth and Hugging Face Jobs! 🚀 This article details how to fine-tune the LFM2.5-1.2B-Instruct model quickly and efficiently. Unsloth boasts approximately 2x faster training and 60% less VRAM usage, making it cost-effective for small models. Small language models are not only affordable but also competitive for targeted tasks, optimized for devices with limited memory. 💻📱 #AI #MachineLearning #Unsloth #HuggingFace #LLM
Source: Hugging Face Blog

「データ不足」の壁を越える:合成ペルソナが日本のAI開発を加速

2026-02-19 15:32
AI has the potential to significantly boost Japan's economy, generating over ¥100 trillion ($650 billion). However, many projects face a critical challenge: a lack of usable training data. Developers building AI systems that understand Japanese language and culture struggle with data scarcity. While English data is abundant, Japan's specific needs for culturally relevant data remain unmet, hindering innovation. #AI #Japan #DataChallenge #Innovation #Technology 🚀📊
Source: Hugging Face Blog

IBM and UC Berkeley Diagnose Why Enterprise Agents Fail Using IT-Bench and MAST

2026-02-18 16:15
IBM and UC Berkeley have teamed up to explore the reasons behind the failures of enterprise agents in IT automation. Their study focuses on tasks like incident triage, log queries, and Kubernetes actions. Traditional benchmarks often indicate failure but don’t explain the underlying reasons. This research aims to enhance our understanding of these systems. #ITAutomation #EnterpriseTech #AIResearch #IBM #UCBerkeley 🤖📊🔍
Source: Hugging Face Blog

One-Shot Any Web App with Gradio's gr.HTML

2026-02-18 00:00
🚀 Gradio 6 introduces a new feature: gr.HTML, which allows for custom templates, scoped CSS, and JavaScript interactivity. This means developers can create any web component easily. Claude, or other advanced LLMs, can generate complete apps with frontend, backend, and state management in one Python file. Various apps, like a Pomodoro Timer, can be built and deployed quickly to Hugging Face Spaces. #Gradio #WebDevelopment #AI #TechUpdates #Python
Source: Hugging Face Blog

NVIDIA Nemotron 2 Nano 9B Japanese: 日本のソブリンAIを支える最先端小規模言語モデル

2026-02-17 23:28
🚀 NVIDIA has unveiled the Nemotron-Nano-9B-v2-Japanese, achieving state-of-the-art performance for models with under 10B parameters. This lightweight model enhances Japanese language understanding and agent capabilities, marking a significant step for enterprise AI development in Japan. The initiative encourages developers to customize models for diverse use cases using the existing Nemotron framework. #NVIDIA #AI #JapaneseLanguage #EnterpriseAI #Innovation
Source: Hugging Face Blog

Custom Kernels for All from Codex and Claude

2026-02-13 00:00
🚀 Exciting advancements in AI coding! Codex and Claude have developed an agent skill that guides coding agents to create production-ready CUDA kernels. They successfully tackled complex tasks, generating functioning kernels for both a diffusers pipeline and a transformers model. This innovation highlights how specialized knowledge can enhance coding efficiency, even in challenging scenarios. #ArtificialIntelligence #Coding #CUDA #MachineLearning #Innovation
Source: Hugging Face Blog

OpenEnv in Practice: Evaluating Tool-Using Agents in Real-World Environments

2026-02-12 00:00
🚀 OpenEnv is a new open-source framework from Meta and Hugging Face, designed to evaluate AI agents in real-world settings. It addresses the common challenges agents face, such as reasoning across multiple steps and interacting with real tools. 📅 Turing has contributed a calendar management environment to help study these agents under realistic constraints, focusing on access control and multi-agent coordination. 🔍 In this article, the authors discuss how OpenEnv operates and why calendars...
Source: Hugging Face Blog

Transformers.js v4 Preview: Now Available on NPM!

2026-02-09 00:00
🚀 Transformers.js v4 (preview) is now available on NPM! After a year of development, users can easily install it with a simple command: `npm i @huggingface/transformers@next`. Regular updates will be published under the next tag until the full release. Stay tuned for more! #TransformersJS #NPM #SoftwareDevelopment #OpenSource #TechNews
Source: Hugging Face Blog

Introducing SyGra Studio

2026-02-05 16:52
🚀 **Introducing SyGra Studio!** SyGra 2.0.0 launches an interactive environment for synthetic data generation. Users can now design flows directly on a visual canvas, simplifying the process without the need for YAML files or terminals. Preview datasets, adjust prompts with inline hints, and monitor live executions, all in one place. This upgrade maintains compatibility with SyGra's existing platform. #SyGraStudio #DataGeneration #Innovation #TechUpdate #AI
Source: Hugging Face Blog

🚀 SyGra V2.0.0

2026-02-05 16:52
🚀 SyGra V2.0.0 has been released, bringing new features and improvements to enhance user experience. This version focuses on optimizing performance and streamlining workflows, making tasks more efficient. Users can expect a smoother interface and updated functionalities to better serve their needs. Stay tuned for further updates! #SyGra #SoftwareUpdate #TechNews
Source: Hugging Face Blog

Nemotron ColEmbed V2: Raising the Bar for Multimodal Retrieval with ViDoRe V3’s Top Model

2026-02-04 15:00
🚀 Exciting advancements in multimodal retrieval are here with the Nemotron ColEmbed V2! This model enhances information retrieval by integrating diverse document types, including text, tables, and images. It builds on foundational vision-language models to unify retrieval processes. Researchers are exploring new multi-vector embedding architectures that improve semantic relationships for better accuracy. Stay tuned for more updates! 📊🔍 #MultimodalRetrieval #AI #MachineLearning #DataScience...
Source: Hugging Face Blog

Community Evals: Because we're done trusting black-box leaderboards over the community

2026-02-04 00:00
🚀 Exciting news in model evaluation! Benchmark datasets on Hugging Face can now host leaderboards where models store their own evaluation scores. This integration allows community members to submit results through pull requests. Verified badges ensure that results are reproducible, enhancing trust in the evaluation process. #MachineLearning #CommunityDriven #HuggingFace #ModelEvaluation #OpenSource
Source: Hugging Face Blog

H Company's new Holo2 model takes the lead in UI Localization

2026-02-03 17:40
🚀 H Company has launched the Holo2-235B-A22B Preview, its most advanced UI localization model yet. This model has achieved impressive scores of 78.5% on Screenspot-Pro and 79.0% on OSWorld G. It focuses on enhancing the localization of UI elements, especially on high-resolution screens. Key to its success is agentic localization, which improves accuracy iteratively, yielding up to 20% gains. Explore the potential of this new model on Hugging Face! 💻✨ #Holo2 #UILocalization #TechInnovation #AI...
Source: Hugging Face Blog

The Future of the Global Open-Source AI Ecosystem: From DeepSeek to AI+

2026-02-03 15:03
The article discusses the evolution of China's open-source AI community since the "DeepSeek Moment" in January 2025. It highlights the strategic shifts and the growth of open artifacts within prominent Chinese AI organizations. For researchers, developers, and policymakers, open-source is positioned as the main approach for future AI developments in China, focusing on sharing resources for large-scale deployment. 🌐🤖 #OpenSourceAI #ArtificialIntelligence #ChinaAI #DeepSeek #Innovation
Source: Hugging Face Blog

Training Design for Text-to-Image Models: Lessons from Ablations

2026-02-03 11:25
🚀 Exciting insights on training text-to-image models are shared in the latest article! The post focuses on the training methods used for their model, PRX, highlighting key strategies for faster training and improved learning outcomes. The team aims to provide a transparent look at their progress and encourages readers to explore their early model version. #TextToImage #AI #MachineLearning #ModelTraining #Innovation
Source: Hugging Face Blog

Introducing NVIDIA Cosmos Policy for Advanced Robot Control

2026-01-29 17:03
🚀 Exciting news from NVIDIA! They are expanding their Cosmos™ world foundation models to address challenges in robotics, autonomous vehicles, and industrial vision AI. The latest development, Cosmos Policy, focuses on enhancing robot control and planning using these advanced models. Stay tuned for more updates! 🤖🌐 #NVIDIA #Robotics #AI #AutonomousVehicles #Innovation
Source: Hugging Face Blog

Introducing Daggr: Chain apps programmatically, inspect visually

2026-01-29 00:00
🚀 Exciting news in AI development! Daggr is a new open-source Python library designed to build AI workflows. It seamlessly connects Gradio apps, ML models, and custom functions. Users can visually inspect intermediate outputs and manage complex pipelines with just a few lines of code. 🖥️ Explore the future of AI application development! #AI #MachineLearning #Python #OpenSource #Daggr
Source: Hugging Face Blog

Nemotron-Personas-Brazil: Co-Designed Data for Sovereign AI

2026-01-28 00:56
Discover how Nemotron is advancing Sovereign AI with a unique approach to Brazilian Portuguese personas. This project focuses on co-designed data that reflects real-world distributions, enhancing AI interactions. By prioritizing local context, the initiative aims to improve AI's understanding and responsiveness in Brazilian settings. #SovereignAI #ArtificialIntelligence #Brazil #DataInnovation 🤖🇧🇷📊
Source: Hugging Face Blog

We Got Claude to Build CUDA Kernels and teach open models!

2026-01-28 00:00
🚀 Exciting advancements in AI! A recent blog post discusses how Claude is being utilized to teach open models to write CUDA kernels. It highlights two approaches: using state-of-the-art models for tough challenges or enhancing smaller models for more complex tasks. The focus is on a new tool called "upskill," which helps generate and assess agent skills. It can lower costs and allow smaller models to tackle difficult, domain-specific problems. #AI #MachineLearning #CUDA #OpenModels #Innovation
Source: Hugging Face Blog

Architectural Choices in China's Open-Source AI Ecosystem: Building Beyond DeepSeek

2026-01-27 15:01
Exploring the evolution of China's open-source AI ecosystem, the latest article delves into architectural and hardware decisions made by companies since the "DeepSeek Moment." Key points include the shift towards openness, diversified models, and the rise of domestic hardware, reflecting competitive strategies in the AI landscape. 🔍💻🇨🇳 #OpenSourceAI #ChinaTech #Innovation #DeepSeek #AIArchitecture
Source: Hugging Face Blog

Alyah ⭐️: Toward Robust Evaluation of Emirati Dialect Capabilities in Arabic LLMs

2026-01-27 10:26
Introducing Alyah ⭐️, a new initiative focused on evaluating Emirati dialect capabilities in Arabic Language Models (LLMs). The project aims to enhance understanding and performance of these models in local dialects, ensuring they meet user needs effectively. Stay tuned for updates on this important development in language technology! 🌍🗣️ #Alyah #ArabicLLMs #EmiratiDialect #LanguageTechnology #AI
Source: Hugging Face Blog

Unlocking Agentic RL Training for GPT-OSS: A Practical Retrospective

2026-01-27 01:53
Unlocking agentic reinforcement learning (RL) enhances traditional LLM training by focusing on multi-step decision-making. This method enables models to optimize performance through direct interaction with their environment, rather than relying on static data. It supports continuous learning by adjusting to outcomes and refining actions over time. Key processes include collecting on-policy data, computing rewards, and updating policies for improved future interactions. #ReinforcementLearning...
Source: Hugging Face Blog

**NVIDIA Earth-2 Open Models Span the Whole Weather Stack**

2026-01-26 14:53
🚀 NVIDIA has launched three new open-source models under the Earth-2 initiative, enhancing weather forecasting capabilities. These models cover various tasks including data assimilation and forecasting. Developers can utilize Earth2Studio and Physics Nemo to build and train their own simulations. Earth-2 offers a flexible platform for customized weather and climate predictions. #NVIDIA #WeatherForecasting #OpenSource #ClimateTech #AI
Source: Hugging Face Blog

AssetOpsBench: Bridging the Gap Between AI Agent Benchmarks and Industrial Reality

2026-01-21 06:25
Introducing AssetOpsBench, a new benchmark and evaluation system for agentic AI! 🤖 This tool focuses on six qualitative dimensions to enhance AI performance in industrial Asset Lifecycle Management. It aims to align AI benchmarks with real-world applications. Explore how AssetOpsBench can transform AI in specific domains! #AI #AssetManagement #Innovation #Technology #Benchmarking
Source: Hugging Face Blog