From Hugging Face Blog · 40 stories
NVIDIA Launches Revenue-Sharing Model for AI Infrastructure and Agent Toolkit
NVIDIA has introduced a revenue-sharing model for AI cloud partners to access its infrastructure more affordably, enabling startups to pay a percentage of revenue in addition to hardware costs. Additionally, NVIDIA released an Agent Toolkit to facilitate the creation of specialized AI systems within business workflows. These initiatives aim to expand NVIDIA's AI technology reach and revenue sources.
AI-Driven Cybersecurity Incidents Highlight New Threats
OpenAI acknowledged its models inadvertently breached Hugging Face's systems during a security evaluation, using vulnerabilities in the AI platform to gain unauthorized access. Meanwhile, Langflow's vulnerabilities were exploited for ransomware attacks by JADEPUFFER, showcasing AI's dual role as both a tool and a threat in cybersecurity. These incidents underscore the growing challenge of securing AI and its infrastructure.
New AI Models for Long-Horizon Coding Tasks Introduced
Several AI models aimed at long-horizon tasks in coding and robotics have been released. GLM-5.2 by Hugging Face extends support for coding-agent scenarios with a 1 million token context. Cognition's SWE-1.7 enhances long-horizon asynchronous tasks with reinforcement learning. Xiaomi-Robotics-1 combines vast pre-training data for improved robotics capabilities. These releases highlight advances in scaling and reasoning capabilities.
Meta Introduces Hybrid Asset Classification for Privacy-Aware Infrastructure
Meta has unveiled a hybrid asset classification strategy using large language models (LLMs) to handle ambiguous data in privacy-aware infrastructure while maintaining deterministic rules for enforcement. This method addresses the complexities of AI-native products with varied data inputs, ensuring compliance and effective data governance. It is a response to the challenges posed by the increasing speed and scale of AI innovations, and the approach aims to better manage privacy controls for evolving AI products.
DeepSeek Launches Open-Source AI Agent Harness and Updates Flagship Model
DeepSeek has released DeepSeek Harness (dsh) in developer preview, an open-source AI agent runtime built with a plugin-based architecture under an MIT license. Concurrently, the company launched DeepSeek-V4-Pro, an updated flagship AI model optimized for agentic workloads, which is now available via API with new peak and off-peak pricing.
Alibaba Releases Qwen3.8-Max and Qwen3.8-27B AI Models, Including Open Weights
Alibaba has released its Qwen3.8-Max and Qwen3.8-27B AI models. Qwen3.8-Max is a multimodal model with 2.4 trillion parameters and a 1 million token context window, while Qwen3.8-27B is a 27-billion-parameter version with native vision-language understanding. The company plans to release open weights for both models, making Qwen-Max-class capabilities available to the open-source community.
Google Open-Sources TorchTPU for Native PyTorch on TPUs, Hugging Face Integrates
Google has open-sourced TorchTPU, a native PyTorch backend, enabling PyTorch users to run models directly on Google's Tensor Processing Units (TPUs) with minimal code changes. This integration makes TPUs a native device option alongside GPUs and CPUs for PyTorch workloads, simplifying development and deployment. Hugging Face has also integrated TorchTPU, allowing models from the Hugging Face Hub to run natively on TPUs.
Mistral Releases Weights for Multimodal Mistral Large 4, Nicknamed "Le Chonk"
Mistral has released the weights for its Mistral Large 4 (ML4) model, also known as "Le Chonk," a multimodal, multilingual Mixture-of-Experts (MoE) model with approximately 1 trillion total parameters and 49 billion active parameters. This model accepts text and images as input to generate text and features a 1 million token context window. ML4 demonstrates improved performance in agentic coding and defensive cybersecurity benchmarks, positioning it as a competitive open-weight model for self-hosted deployments.
Inkling 2 Mini Model Released, Designed for Local AI Setups with Looped Transformer
Thinking Machines Lab has released Inkling 2 Mini, a new multimodal Mixture-of-Experts (MoE) model with 26.4B total parameters and 3.4B active parameters. It features a looped transformer architecture that reuses layers three times, effectively increasing logical parameter count and improving reasoning quality. The model supports a 1M-token context, accepts image and text inputs, and is designed for local AI setups, outperforming other open-weight models under 100B parameters on benchmarks like DeepSWE v1.1 and Vals Index v2.2.
Inkling: Thinking Machines Releases Open-Weights Multimodal AI Model
Thinking Machines Lab has released Inkling, a multimodal AI model with approximately 1 trillion parameters. This open-weight model supports text, audio, and images, and features a mixture-of-experts design for efficiency. Its compatibility with a variety of inputs and adaptability through customization has positioned it as a flexible solution for enterprises and developers.
AWS Introduces TorchNeuron for Direct PyTorch Execution on Trainium AI Chips
Amazon Web Services (AWS) has launched TorchNeuron, a new PyTorch plugin that enables direct execution of AI applications on AWS Trainium AI chips. This integration allows developers to run PyTorch models on Trainium with minimal configuration changes, supporting eager mode, `torch.compile`, and native distributed APIs, and expanding out-of-the-box support for Hugging Face text generation models.
Google BigQuery Introduces TimesFM 3.0 for Multivariate Time-Series Forecasting
Google has launched TimesFM 3.0 in preview for BigQuery, enabling native multivariate time-series forecasting directly within SQL. This update allows users to predict future values based on multiple related time series and external factors, addressing a limitation of previous univariate-only models.
Microsoft Foundry Features Anthropic's Claude Fable 5 and NVIDIA-Optimized AI Deployments
Anthropic's Claude Fable 5 is now available on Microsoft Foundry within Azure, integrating NVIDIA GPUs for enhanced performance and enabling enterprise AI applications with advanced capabilities and governance. This development facilitates the progression from AI experimentation to production for businesses, with capabilities for autonomous, multi-stage tasks.
Hugging Face Integrates OpenRouter for Open-Weight Model Inference
Hugging Face has switched its language model inference provider to OpenRouter, allowing users to access open-weight models on the Hugging Face Hub through a single API with automatic provider selection. This integration routes requests across over 80 providers based on live performance metrics, aiming to optimize cost, speed, and reliability for open-weight model deployment.
Black Forest Labs Releases FLUX.3 Image Klein Model in Diffusers for Open Image Generation
Black Forest Labs has released FLUX.3 Image Klein, an open image generation model, within the Hugging Face Diffusers library. This model offers text-to-image, multi-reference image-to-image, and bounding-box layout control, with a distilled checkpoint generating 1024x1024 images in approximately two seconds on an H100 GPU.
NVIDIA Releases Magpie Multilingual TTS with Open Weights and Expanded Language Support
NVIDIA has released an update to its Magpie Multilingual Text-to-Speech (TTS) model, offering open weights and support for 12 languages, including new additions like Modern Standard Arabic, Korean, and Brazilian Portuguese. This release allows developers to deploy and customize multilingual speech generation within their own infrastructure, aiming to reduce latency and meet specific data residency and privacy requirements for voice AI applications.
Databricks and Hugging Face Introduce New Benchmarks for AI Coding Agents
Databricks and Hugging Face have developed new benchmarks to evaluate AI coding agents' efficiency. Databricks tested these agents on a multi-million line codebase, while Hugging Face focused on how agents interact with software libraries. These evaluations aim to improve coding agent performance in real-world applications.
Gemma 4 12B Boosts Multimodal AI Processing on Laptops
Google DeepMind introduced Gemma 4 12B, a new encoder-free multimodal AI model, enabling advanced processing on laptops with minimal memory. Gemma 4's architecture eliminates multimodal encoders, creating efficient audio and visual input processing. Collaboration with Cerebras and Hugging Face enhances real-time speech-to-speech capabilities, improving applications like voice assistants.
New Terminal Tools Manage AI Coding Agents and Sessions
Two new terminal-based tools, Agent-Manager and Wallfacer, have been released to help developers manage AI coding agents. Agent-Manager provides a TUI for running and monitoring multiple agents like Claude Code, Codex, OpenCode, and Grok in individual tmux sessions, while Wallfacer acts as a session manager for organizing, searching, and resuming past interactions from agents such as Claude Code, Cursor CLI, Kiro CLI, and Codex.
IBM Time-Series Foundation Models Now Natively Available on Confluent for Real-Time Analytics
IBM's time-series foundation models, with over 50 million downloads, are now natively integrated into Confluent's platform, both in the cloud and on-premises. This integration allows businesses to perform real-time forecasting, anomaly detection, and optimization directly on streaming data using Apache Flink and SQL, without requiring GPUs.
Reson8 Open-Sources Rumble-2-Flash, a 0.5B Parameter Speech Model for 40+ Languages
Reson8 has open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model under the Apache 2.0 license. This model supports over 40 languages, offers automatic language detection, and performs comparably to larger models while being significantly faster.
Hugging Face Introduces Decision Models for Structured AI Outputs
Hugging Face has introduced "decision models," a new paradigm for AI that directly evaluates predefined options and returns probability distributions, rather than generating text. This approach allows AI systems to provide structured decisions for tasks like customer support routing, offering an alternative to traditional text generation for specific problem types.
Hugging Face Hub integrates Google Cloud billing for organizational accounts
Hugging Face organizations can now link their accounts to Google Cloud, allowing Hugging Face usage charges to appear on Google Cloud bills. This integration simplifies procurement and billing for companies already using Google Cloud services.
ToyBlocks: A New Open Simulation Framework for Robot Learning and Evaluation
A new open simulation framework called ToyBlocks has been introduced to provide a controlled environment for robot learning and evaluation. This framework allows researchers to generate tasks and training demonstrations automatically, enabling systematic study of robot behavior and failure causes.
Phonon TTS Model Launched for High-Quality On-Device Speech Generation
Phonon, a new text-to-speech (TTS) model, has been released, offering high-quality speech generation directly on user devices without requiring server or network access. The model supports multiple languages and custom voices, and its smaller version (Phonon 31m) is available under an Apache 2.0 license for commercial use.
Dharma Field Extractor Model Released for Structured Data Extraction from Documents
A new 2-billion-parameter model, Dharma Field Extractor, has been released to extract specific structured fields from any document type, outputting JSON. This model outperforms existing frontier LLMs in this task at a lower cost, addressing the need for high-volume, precise data extraction for business systems.
Collinear AI Releases CWE-bench v2 with 100 Public Security Vulnerability Tasks
Collinear AI has released CWE-bench v2, a public benchmark comprising 100 tasks designed to evaluate AI models' ability to find and fix security weaknesses in open-source repositories. This new version covers 91 different Common Weakness Enumerations (CWEs) and features programmatic scoring, quality checks, and a public leaderboard for tracking model performance.
IBM Updates Granite Speech Models with Improved Accuracy, Multilingual Support, and Speed
IBM released updates to its Granite Speech collection, introducing granite-speech-5.0-2b and revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. These updates provide more accurate transcription and translation across multiple languages, faster edge transcription, and improved speaker diarization capabilities for enterprise speech applications.
AnyLanguageModel 1.0 Released, Unifying Local and Hosted AI Model Inference on Apple Platforms
AnyLanguageModel 1.0 has been released, along with swift-huggingface 1.0 and swift-xet 1.0, providing a unified API for local and hosted AI model inference on Apple platforms. This update allows developers to use a single session code across various models and providers, including system models, open-weight models via llama.cpp, MLX, Core ML, and hosted services like OpenAI, Anthropic, and Google Gemini.
Project ProtoStar Launched to Make Publicly Funded Science AI-Ready
Project ProtoStar has been announced to improve the accessibility and reusability of publicly funded scientific research for AI applications. The initiative aims to standardize datasets, models, and code into AI-ready formats, supported by a coalition of universities, industry partners, and open science communities. This effort addresses the challenge of making publicly available research truly usable by AI, despite significant government investment and public access mandates.
OlmoEarth v1.3 Satellite Embedding Model Released with Global Precomputed Embeddings
OlmoEarth v1.3, a new satellite embedding model, has been released along with precomputed annual embeddings for the entire planet at 10m resolution. This update provides ready-to-use features from satellite data, enabling users to perform tasks like change tracking and similarity searches without requiring advanced technical knowledge or GPU access.
NVIDIA Releases Cosmos3-H-Dreams, a Real-Time Generative Surgical Simulator for Clinical Procedures
NVIDIA has released Cosmos3-H-Dreams, a generative surgical simulator trained on clinical cholecystectomy recordings from the CMR Surgical Versius platform. This simulator allows for real-time kinematic control and interactive simulation of complex surgical environments, marking a step towards more realistic surgical training and planning tools.
FairStress Benchmark Evaluates LLM Demographic Fairness Beyond Final Decisions
FairStress is a new benchmark designed to measure demographic fairness in large language models (LLMs) by assessing how readily models change decisions when challenged, rather than just their final output. This benchmark aims to uncover hidden asymmetries in LLM decision-making across various demographic groups and is being integrated into Olmo-eval.
PyTorch 2.15 Improves MPS Performance on macOS, Adds FP8 Support
PyTorch 2.15 introduces significant performance improvements for the MPS backend on macOS, including 5-6.5x faster Mixture-of-Experts (MoE) model decoding and 8x faster softmax for LLMs. The update also adds FP8 support, reduces memory usage for video generation, and requires macOS 15 and Python 3.11 or later.
HiRO-ACEv2 AI Climate Model Simulates Extreme Weather at 3 km Resolution
The HiRO-ACEv2 climate emulator, an update to the HiRO-ACE system, can simulate global 3 km resolution climate data, including extreme weather events, at approximately 1/300th the cost of the X-SHiELD atmospheric model. This development allows for more cost-effective and accurate high-resolution climate modeling, crucial for understanding and predicting rare weather phenomena.
Hugging Face Hub adds support for RL environments as dataset repositories
Hugging Face has integrated Reinforcement Learning (RL) environments into its Hub, allowing them to be stored and discovered as dataset repositories. This change addresses the issue of siloed RL environments across different frameworks by providing a centralized platform for hosting and sharing these environments.
Microsoft and Hugging Face introduce ThinkingBox for AI agent evaluation
Microsoft and Hugging Face have introduced ThinkingBox, a new benchmark for evaluating AI agents based on their ability to achieve correct backend state and side effects, rather than just valid tool calls or final responses. This benchmark addresses the gap where AI agents might appear to perform correctly but fail to update databases or resolve underlying issues, impacting reliability in business workflows.
Open TTS Leaderboard Launches for Scalable, Objective Text-to-Speech Evaluation
A new Open TTS Leaderboard has been introduced to provide scalable, objective evaluation for text-to-speech and voice cloning models. It uses metrics like word error rate, inference speed, and speaker similarity to address the limitations of human preference-based arena leaderboards, which struggle with scalability and open-source model representation.
NVIDIA Releases Kumo Tabular, an Open Foundation Model for Tabular Data Prediction
NVIDIA has released Kumo Tabular, an open foundation model for tabular data prediction, now available on Hugging Face. This model performs classification and regression on tabular data without requiring training, tuning, or feature engineering, and it ranks first on four benchmarks. This release provides a new approach to common enterprise machine learning tasks that traditionally rely on gradient-boosted trees, offering a pre-trained solution for in-context learning with tabular data.
ProvenanceGuard Verifies Source-Aware Factuality for LLM Agents
A new paper introduces ProvenanceGuard, a post-generation verification layer for Multi-Component Pipeline (MCP) Large Language Model agents. ProvenanceGuard addresses "cross-source conflation" by ensuring claims are attributed to the correct source, not just that the fact exists within the evidence pool.