✷ The archive · 11 labs · 871 dispatches
The labs
No more opening a dozen official blogs every morning. First-hand releases from OpenAI, Anthropic, DeepMind and the rest, each with its substance pulled out.
HELMET: Holistically Evaluating Long-context Language Models
Hugging Face and Princeton researchers introduced HELMET, a comprehensive benchmark designed to replace synthetic tests like needle-in-a-haystack with diverse, controllable, and reliable real-world evaluations for long-context language models.
Cohere Integration with Hugging Face Inference Providers
Hugging Face has added Cohere as a supported Inference Provider, allowing users to run serverless inference for a wide range of Cohere and Cohere Labs models directly on the Hub.
Gradio Framework Overview: Beyond UI Library Capabilities
Hugging Face details how Gradio has evolved into a comprehensive AI-focused framework providing universal API access, server-side rendering, and specialized ML resource management.
Hugging Face and Protect AI Security Partnership: 6-Month Progress Report
Hugging Face and Protect AI have scanned 4.47 million model versions across 1.41 million repositories to identify 352,000 unsafe or suspicious issues, enhancing open-source AI security through the Guardian scanning technology.
Hugging Face acquires Pollen Robotics to expand open-source robotics hardware
Hugging Face has acquired Pollen Robotics to integrate open-source humanoid hardware with the LeRobot software ecosystem, starting with the sale of the Reachy 2 robot.
Visual Salamandra 7B Release
Hugging Face's Language Technologies Lab has released Visual Salamandra, a 7-billion parameter multimodal model that extends the Salamandra LLM to support images and video with a focus on European linguistic diversity.
Hugging Face and Cloudflare FastRTC Integration
Hugging Face and Cloudflare have partnered to provide FastRTC developers with free access to Cloudflare's global TURN server network, simplifying the deployment of low-latency real-time audio and video AI applications.
Arabic Leaderboards: Arabic Instruction Following and AraGen Updates
Hugging Face and Inception announce the launch of the Arabic-Leaderboards Space, featuring the first public Arabic Instruction Following (Arabic IFEval) benchmark and an updated AraGen-03-25 generative leaderboard.
Llama 4 Maverick & Scout Release Notes
Meta has released Llama 4 Maverick and Llama 4 Scout, natively multimodal Mixture-of-Experts (MoE) models featuring active parameters of 17B and context windows up to 10M tokens.
Gradio 1 Million Users Milestone and Development Philosophy
Hugging Face's Gradio has reached over 1 million monthly developers, achieving growth by prioritizing low-level primitives over high-level abstractions and focusing specifically on the machine learning niche.
Hugging Face NLP Course transitions to LLM Course
Hugging Face is rebranding and expanding its NLP Course into the LLM Course to incorporate modern Large Language Model research, fine-tuning, and reasoning models while maintaining classic NLP foundations.
Efficient Request Queueing for LLM Performance Optimization
TNG Technology Consulting GmbH outlines a strategy for optimizing LLM performance by implementing fair scheduling and metric-based backpressure in an upstream LLM-Server to prevent power users from blocking others.
Hugging Face Secrets Management Scaling with Infisical
Hugging Face migrated to Infisical to centralize secrets management across a multi-cloud environment, eliminating secret sprawl and improving security through automated Kubernetes integration and RBAC.
Text Generation Inference (TGI) Intel Gaudi Integration
Hugging Face has integrated Intel Gaudi hardware support directly into the Text Generation Inference (TGI) main codebase, providing a production-ready serving solution for LLMs on Intel AI accelerators.
DeepSeek-V3-0324 Release Notes
DeepSeek has released DeepSeek-V3-0324, an updated version of the base model for R1 that features improved instruction following, coding, and math capabilities under an MIT license.
Training and Finetuning Reranker Models with Sentence Transformers
Hugging Face provides a comprehensive guide and framework via Sentence Transformers for training and finetuning Cross Encoder reranker models to optimize domain-specific retrieval performance.
Gradio Dataframe Update: New Features and Enhancements
Hugging Face has updated the gr.Dataframe component in Gradio, introducing multi-cell selection, column pinning, search and filter functionality, and improved accessibility.
Hugging Face Inference Endpoints Analytics Update
Hugging Face has updated its Inference Endpoints analytics dashboard with real-time metrics, customizable time ranges, and a detailed replica lifecycle view to improve monitoring and debugging.
Open R1: Running OlympicCoder 7B Locally for Coding
Hugging Face provides a guide on deploying the OlympicCoder 7B model locally using LM Studio and the Continue VS Code extension to achieve competitive coding performance that rivals Claude 3.7 Sonnet and GPT-4o on LiveCodeBench.
Hugging Face Response to White House AI Action Plan RFI
Hugging Face advocates for the fundamental role of open source and open science in AI development, arguing that open models are increasingly matching or surpassing closed commercial systems in performance and efficiency.
NVIDIA GTC 2025 Physical AI Releases: Cosmos Transfer, Physical AI Dataset, and Isaac GR00T N1
NVIDIA has released Cosmos Transfer for controllable world scene generation, a 15TB Physical AI Dataset, and Isaac GR00T N1, the first open foundation model for general humanoid reasoning.
Hugging Face Xet Storage Integration
Hugging Face has begun migrating repositories from LFS to Xet storage, utilizing content-defined chunking to significantly reduce upload and download times for massive AI models and datasets.
Gemma 3 Release Notes / What's New
Google has released Gemma 3, a multimodal, multilingual open-weight LLM family ranging from 1B to 27B parameters with context windows up to 128k tokens.
Open R1 Update #3: OlympicCoder and Code Reasoning Insights
Hugging Face introduces OlympicCoder, a set of code reasoning models that outperform frontier models on IOI problems, alongside new datasets and technical lessons for training reasoning models.
LeRobot L2D: World's Largest Open-Source Self-Driving Dataset
Hugging Face and Yaak have released Learning to Drive (L2D), a multimodal self-driving dataset featuring over 5,000 hours of driving data from 30 German cities to enable end-to-end spatial intelligence training.
LLM Inference on Edge: Running Local LLMs via React Native
Hugging Face provides a comprehensive guide to building a privacy-focused mobile application using React Native and llama.rn to run quantized GGUF models locally on Android and iOS.
Aya Vision: Advancing Multilingual Multimodality with 8B and 32B Models
Cohere For AI has released Aya Vision, a family of open-weight vision-language models (8B and 32B) supporting 23 languages, outperforming larger models in multilingual multimodal tasks.
Hugging Face and JFrog Partnership for Enhanced AI Model Security
Hugging Face has integrated JFrog's scanner into the Hugging Face Hub to reduce false positives and detect malicious code within model weights across various serialization formats.
Tracing and Evaluating smolagents with Arize Phoenix
Hugging Face demonstrates how to use Arize Phoenix with smolagents to implement real-time tracing and LLM-as-a-judge evaluations for agentic workflows.
Hugging Face and IISc Partner to Open-Source the Vaani Dataset
Hugging Face has partnered with the Indian Institute of Science (IISc) and ARTPARK to provide global access to Vaani, a massive multi-modal, multi-lingual dataset designed to represent India's linguistic diversity.
FastRTC: The Real-Time Communication Library for Python
Hugging Face has released FastRTC, a Python library designed to simplify the development of real-time audio and video AI applications by handling the WebRTC and WebSocket communication layers.
Hugging Face Remote VAEs for Inference Endpoints
Hugging Face has introduced an experimental feature to delegate the VAE decoding process to remote endpoints, reducing VRAM requirements for high-resolution image and video synthesis on consumer GPUs.
SigLIP 2 release notes / what's new
Google has released SigLIP 2, a family of multilingual vision-language encoders that outperform the original SigLIP across all scales in zero-shot classification, image-text retrieval, and VLM transfer performance.
SmolVLM2: Bringing Video Understanding to Every Device
Hugging Face has released SmolVLM2, a family of efficient vision and video language models in 2.2B, 500M, and 256M parameter sizes designed to enable local video understanding on devices ranging from phones to servers.
PaliGemma 2 Mix Release Notes
Google has released PaliGemma 2 mix, a family of vision language models fine-tuned on a diverse mix of tasks including OCR, captioning, and object detection to demonstrate the potential of PaliGemma 2 pre-trained checkpoints.
Hugging Face adds Hyperbolic, Nebius AI Studio, and Novita as Serverless Inference Providers
Hugging Face has integrated Hyperbolic, Nebius AI Studio, and Novita as serverless inference providers, expanding access to models like DeepSeek-R1 and FLUX.1 directly via the Hub and client SDKs.
Hugging Face Open LLM Leaderboard Update: Integrating Math-Verify for Improved Math Evaluation
Hugging Face has re-evaluated 3,751 models on the Open LLM Leaderboard using Math-Verify to fix parsing errors and format strictness, resulting in a significant reshuffling of the MATH-Hard rankings.
Hugging Face Integrates Fireworks.ai as an Inference Provider
Hugging Face has added Fireworks.ai as a supported Inference Provider on the Hub, enabling serverless inference for models like DeepSeek-R1 and Llama-3.2-90B-Vision-Instruct across the HF ecosystem.
Hugging Face: Optimizing Cost and Latency for 1 Billion Classifications
Hugging Face provides a framework and benchmarks for reducing the cost of large-scale encoder model inference, demonstrating that NVIDIA L4 GPUs and optimized batch sizes can process 1 billion text classifications for as little as $253.82.
Hugging Face Video Dataset Scripts
Hugging Face has introduced a set of open video dataset scripts designed to simplify the creation of high-quality, filtered datasets for fine-tuning video generation models.
Hugging Face Xet-backed Repositories: Accelerating Hub Transfers with Block-Level Aggregation
Hugging Face is introducing a chunk-based deduplication system using xet-core and hf_xet to accelerate uploads and downloads by 2-3x through block-level aggregation.
Open R1 Update #2: OpenR1-Math-220k Dataset and Reasoning Insights
Hugging Face introduces OpenR1-Math-220k, a large-scale math reasoning dataset designed to reconstruct DeepSeek R1's distillation pipeline, while sharing community insights on GRPO and Chain-of-Thought length control.
The Open Arabic LLM Leaderboard 2
Hugging Face and partners have released the Open Arabic LLM Leaderboard 2 (OALL v2), replacing saturated and machine-translated benchmarks with native Arabic datasets and a new RAG evaluation framework to more accurately measure Arabic language capabilities.
DABStep: Data Agent Benchmark for Multi-step Reasoning
Hugging Face and Adyen have introduced DABStep, a benchmark of over 450 real-world data analysis tasks that reveals current state-of-the-art AI agents achieve only 16% accuracy on complex reasoning tasks.
pi0 and pi0-FAST: Vision-Language-Action Models for General Robot Control
Hugging Face has integrated pi0 and pi0-FAST, generalist Vision-Language-Action (VLA) models developed by Physical Intelligence, into the LeRobot repository to enable versatile robot control across diverse embodiments.
Hugging Face Open-source DeepResearch
Hugging Face has open-sourced a framework to reproduce OpenAI's Deep Research capabilities, achieving a 55.15% score on the GAIA benchmark using a code-native agentic approach.
Open-R1 Update #1: Replicating DeepSeek-R1 Training and Synthetic Data
Hugging Face provides a first progress update on the Open-R1 project, detailing the reproduction of DeepSeek-R1 evaluation scores, the integration of GRPO into TRL, and strategies for scaling synthetic reasoning data generation.
Mini-R1: Reproducing DeepSeek-R1 Reasoning via GRPO and the Countdown Game
Hugging Face demonstrates how to reproduce the DeepSeek-R1 "aha moment" of self-correction and reasoning by training a Qwen2.5-3B model using Group Relative Policy Optimization (GRPO) on the Countdown Game.
Hugging Face AI Tools for Art Newsletter Issue 1
Hugging Face has launched a monthly newsletter detailing the state of open-source creative AI, highlighting 2024's shift to Diffusion Transformers and the emergence of high-quality open video and audio models.
How to Deploy and Fine‑Tune DeepSeek R1 Models on AWS
Hugging Face shows how to deploy and fine‑tune DeepSeek R1 and its distilled variants on AWS using Hugging Face Inference Endpoints, Amazon Bedrock Marketplace, Amazon SageMaker AI (GPU and Neuron instances), and EC2 Neuron with the Hugging Face Neuron Deep Learning AMI.