✷ 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.
DeepInfra Integration with Hugging Face Inference Providers
Hugging Face has added DeepInfra as a supported Inference Provider, enabling serverless access to over 100 models, including DeepSeek V4 and Kimi-K2.6, directly through the Hub and client SDKs.
NVIDIA Nemotron 3 Nano Omni release notes / what's new
NVIDIA has released Nemotron 3 Nano Omni, an omni-modal model capable of long-context reasoning across text, images, video, and audio, delivering best-in-class accuracy on document intelligence and video understanding benchmarks.
OpenAI Privacy Filter: Building Scalable PII Detection Web Apps
OpenAI has released Privacy Filter, an open-source 1.5B-parameter PII detector capable of labeling eight categories of sensitive data across a 128k context window.
DeepSeek-V4 Release Notes: Efficient 1M-Token Context for AI Agents
DeepSeek-V4 introduces a 1M-token context window powered by a hybrid CSA/HCA attention mechanism, specifically optimized for long-running agentic workloads and tool-use trajectories.
Using Transformers.js in a Chrome Extension
Hugging Face provides a technical guide on integrating Transformers.js into a Chrome Extension using Manifest V3, featuring a background-hosted model architecture powered by Gemma 4 E2B.
QIMMA: A Quality-First Arabic LLM Leaderboard
Hugging Face and partners introduced QIMMA, a new Arabic LLM leaderboard that implements a rigorous quality validation pipeline to ensure benchmarks reflect genuine language capability.
Hugging Face: AI and the Future of Cybersecurity
Hugging Face argues that open-source AI models and tooling are critical for cybersecurity defense to counter the risks posed by autonomous vulnerability-finding systems like Mythos.
Hugging Face transformers-to-mlx Skill and Test Harness
Hugging Face has released a Skill and a non-agentic test harness to streamline the porting of language models from the transformers library to mlx-lm while maintaining high code quality and reviewer signal.
Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers
Hugging Face published a guide on training and finetuning multimodal embedding and reranker models with Sentence Transformers, demonstrating how task-specific finetuning improves performance on retrieval tasks such as Visual Document Retrieval.
Ecom-RLVE: Adaptive Verifiable Environments for E-Commerce Conversational Agents
Hugging Face introduces EcomRLVE-GYM, a framework for training e-commerce agents using eight verifiable, multi-turn environments with adaptive difficulty scaling to bridge the gap between conversational fluency and actual task completion.
VAKRA Benchmark Analysis: Agent Reasoning, Tool Use, and Failure Modes
Hugging Face announced the VAKRA benchmark, a tool‑grounded, executable suite that evaluates AI agents on compositional reasoning across 8,000+ APIs and document sources, revealing widespread failures in tool selection, multi‑hop reasoning, and policy adherence.
HCompany HoloTab Release
HCompany has released HoloTab, a Chrome extension powered by the Holo3 model that allows users to automate web tasks through natural language descriptions or recorded routines.
Sentence Transformers v5.4: Multimodal Embedding and Reranker Models
Sentence Transformers v5.4 introduces multimodal embedding and reranker models for encoding and comparing text, images, audio, and video.
Waypoint-1.5 Release Notes
Hugging Face and Overworld have released Waypoint-1.5, a real-time video world model that enables high-fidelity interactive generative environments to run locally on consumer GPUs.
Safetensors Joins the PyTorch Foundation
Safetensors has transitioned to a foundation-hosted project under the PyTorch Foundation and Linux Foundation to ensure vendor-neutral governance and community-driven development.
Gemma 4 release: open‑source multimodal models with on‑device support
Gemma 4, a new open‑source multimodal family from Google DeepMind, is released on Hugging Face with image, audio, video support, up to 256 K context, and day‑0 compatibility with transformers, llama.cpp, MLX, Rust, and WebGPU.
Falcon Perception and Falcon OCR Release
Hugging Face and TII introduce Falcon Perception, a 0.6B-parameter early-fusion Transformer for open-vocabulary grounding, and Falcon OCR, a 0.3B-parameter high-throughput document understanding model.
Gradio Server: Integrating Custom Frontends with Gradio Backend
Hugging Face introduces gradio.Server, a FastAPI extension that allows developers to use any custom frontend framework while retaining Gradio's queuing, API infrastructure, and ZeroGPU support.
Granite 4.0 3B Vision release notes / what's new
IBM has released Granite 4.0 3B Vision, a compact multimodal model optimized for enterprise document understanding, featuring high-accuracy table extraction, chart reasoning, and key-value pair extraction.
OpenMed CodonRoBERTa multi-species mRNA language models release
OpenMed released an end-to-end protein engineering pipeline with CodonRoBERTa-large-v2 (perplexity 4.10, CAI 0.404) and a 25-species codon‑optimization model suite trained in 55 GPU‑hours for $165.
TRL v1.0 release notes / what's new
Hugging Face releases TRL v1.0, a stable post-training library implementing over 75 methods, featuring a dual-track stability model to balance rapid experimental iteration with production-grade reliability.
Hugging Face OpenClaw Migration Guide
Hugging Face provides two methods—Inference Providers and local llama.cpp setup—to migrate OpenClaw agents from restricted Claude models to open-source alternatives.
EVA End-to-End Evaluation Framework for Voice Agents
EVA is a new end‑to‑end framework that jointly evaluates voice agents on accuracy and conversational experience, revealing a consistent trade‑off between task success and user satisfaction.
Domain-Specific Embedding Fine-Tuning with NVIDIA Nemotron – Under a Day
NVIDIA and Hugging Face released a single‑GPU, under‑a‑day pipeline that fine‑tunes the Llama‑Nemotron‑Embed‑1B‑v2 model on synthetic domain data, delivering >10% retrieval gains and up to 26% improvement on real enterprise datasets.
State of Open Source on Hugging Face: Spring 2026
Hugging Face reports a massive expansion of the open source AI ecosystem in 2025, characterized by China surpassing the U.S. in model downloads and the rapid emergence of robotics as the largest dataset category.
Holotron-12B High Throughput Computer Use Agent
H Company released Holotron-12B, a multimodal computer-use model based on NVIDIA Nemotron-Nano-2 VL that uses a hybrid SSM-Attention architecture to achieve high inference throughput for agentic workloads.
Hugging Face Storage Buckets Release
Hugging Face has introduced Storage Buckets, a mutable, S3-like object storage system backed by Xet for efficient handling of intermediate ML artifacts like checkpoints and processed data.
Keep the Tokens Flowing: Lessons from 16 Open-Source RL Libraries
Hugging Face surveyed 16 open-source RL libraries and found that async RL training separates inference and training onto different GPU pools, uses a rollout buffer, and pushes weights asynchronously, with Ray dominating orchestration and NCCL broadcast the common weight‑sync method.
LeRobot v0.5.0 release notes / what's new
Hugging Face has released LeRobot v0.5.0, introducing full Unitree G1 humanoid support, new VLA policies like Pi0-FAST and Wall-X, and significant dataset performance optimizations.
Ulysses Sequence Parallelism for Million-Token Context Training
Hugging Face has integrated Ulysses Sequence Parallelism into Accelerate, Transformers, and TRL, enabling the training of LLMs with million-token contexts by distributing attention computation across multiple GPUs.
Bringing Robotics AI to Embedded Platforms: Dataset Recording, VLA Fine-Tuning, and On-Device Optimizations
Hugging Face and NXP provide a technical guide on deploying Vision-Language-Action (VLA) models on the i.MX 95 SoC, emphasizing dataset consistency, architectural decomposition, and asynchronous inference to achieve real-time robotic control.
Hugging Face Modular Diffusers Release
Hugging Face has introduced Modular Diffusers, a composable framework that allows users to build diffusion pipelines by mixing and matching reusable blocks rather than writing entire pipelines from scratch.
PRX Part 3: Training a Text-to-Image Model in 24 Hours
Hugging Face and Photoroom demonstrate a text-to-image model trained in 24 hours using 32 H200 GPUs on a $1500 budget, combining pixel-space training, token routing, and representation alignment.
Mixture of Experts (MoEs) in Transformers
Hugging Face has redesigned the transformers library to make Mixture of Experts (MoEs) first-class citizens through a new weight loading refactor, a pluggable expert backend, and native expert parallelism.
Train AI models with Unsloth and Hugging Face Jobs
Hugging Face has integrated Unsloth with Hugging Face Jobs to enable fast, low-cost LLM fine-tuning, specifically optimized for small models like LiquidAI/LFM2.5-1.2B-Instruct.
GGML and llama.cpp join Hugging Face
GGML, the creators of llama.cpp, have joined Hugging Face to provide sustainable resources for local AI inference and streamline the integration between the Transformers library and local model deployment.
IBM and UC Berkeley Diagnose Enterprise Agent Failures Using IT-Bench and MAST
IBM Research and UC Berkeley introduced MAST (Multi-Agent System Failure Taxonomy) to diagnose why enterprise IT agents fail, revealing that frontier models suffer from isolated verification errors while open models face cascading systemic collapses.
Gradio 6 gr.HTML: One-Shot Web App Development
Gradio 6 introduces enhanced gr.HTML support for custom templates, scoped CSS, and JavaScript interactivity, enabling the creation of complex web components within a single Python file.
Hugging Face CUDA Kernels Agent Skill
Hugging Face has introduced an agent skill that enables coding agents like Claude and Codex to write, benchmark, and integrate production-ready CUDA kernels for transformers and diffusers libraries.
OpenEnv: Evaluating Tool-Using Agents in Real-World Environments
Hugging Face and Meta introduce OpenEnv, an open-source framework that evaluates AI agents against real systems and production-grade environments like the Calendar Gym to bridge the gap between research and production reliability.
Transformers.js v4 release notes / what's new
Hugging Face has released Transformers.js v4, introducing a new C++ rewritten WebGPU runtime for hardware acceleration across browsers and server-side runtimes, alongside a standalone tokenizers library.
SyGra 2.0.0 Studio Release
SyGra 2.0.0 introduces Studio, a visual interactive environment for designing and executing synthetic data generation workflows without needing to manually edit YAML files.
Hugging Face Community Evals
Hugging Face has introduced Community Evals, a decentralized system for reporting and aggregating model benchmark scores directly on the Hub to increase transparency and reproducibility.
H Company Holo2-235B-A22B Preview Release
H Company has released Holo2-235B-A22B Preview, a UI localization model that achieves state-of-the-art performance on Screenspot-Pro and OSWorld G benchmarks.
The Future of the Global Open-Source AI Ecosystem: From DeepSeek to AI+
Hugging Face analyzes how open source has become the dominant strategy for Chinese AI organizations, shifting from isolated model breakthroughs to a scalable, integrated ecosystem of models, hardware, and infrastructure.
Training Design for Text-to-Image Models: Lessons from Ablations
The provided source material for the Hugging Face post on text-to-image model training design is unavailable due to a 429 Too Many Requests error.
Introducing Daggr: Chain AI Apps Programmatically with Visual Inspection
Hugging Face has released Daggr, an open-source Python library that allows developers to programmatically chain Gradio apps, ML models, and custom functions into workflows with an automatically generated visual canvas for debugging and state management.
Hugging Face Upskill: Transferring Expert Capabilities to Smaller Models via Agent Skills
Hugging Face introduced upskill, a tool that uses high-capability models like Claude Opus 4.5 to generate validated 'agent skills' that improve the performance and token efficiency of smaller or open-source models on complex tasks such as CUDA kernel development.
Architectural Choices in China's Open‑Source AI Ecosystem: From DeepSeek R1 to a Hardware‑First, MoE‑Driven Landscape
One year after DeepSeek R1’s open‑source release, China’s AI community shifted from chasing the biggest single‑model performance to building flexible, cost‑effective, and hardware‑aware AI systems. Mixture‑of‑Experts (MoE) became the default architecture, enabling huge models to run affordably by activating only a subset of experts per request. Multimodal races exploded, with open releases for text‑to‑image, video, audio, 3‑D, and agents, each bundled with full toolchains. Small models (≤30 B) surged in popularity for local deployment and fine‑tuning, while large MoE models serve as teacher nets for distillation. Apache 2.0 and MIT licenses now dominate, removing legal friction and accelerating commercial adoption. A hardware‑first mindset emerged: releases ship with quantization, inference, and serving stacks tuned for domestic chips (Huawei Ascend, Cambricon, Kunlun), and training pipelines are openly documented. The competitive edge now lies in system design, deployment efficiency, and open‑source ecosystem integration rather than raw model size.
Alyah: Emirati Dialect Benchmark for Arabic LLMs
Hugging Face and partners introduced Alyah, a manually curated benchmark of 1,173 samples designed to evaluate the linguistic and cultural capabilities of Arabic LLMs in the Emirati dialect.