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.

251

Building the Hugging Face MCP Server

Hugging Face has released an official Model Context Protocol (MCP) server that allows AI assistants to dynamically access the Hugging Face Hub and thousands of Gradio-based AI applications via a single URL.

252

ScreenEnv Release: Deploying Full Stack Desktop Agents

Hugging Face has released ScreenEnv, a Python library that enables the creation of isolated Ubuntu desktop environments in Docker containers for testing and deploying GUI agents.

253

Hugging Face Asynchronous Robot Inference

Hugging Face introduces asynchronous robot inference to decouple action prediction from execution, reducing robot idleness and achieving up to a 2x speedup in task completion time.

254

Hugging Face Gradio MCP Servers Integration

Hugging Face has integrated the Model Context Protocol (MCP) into Gradio (v5.28.0), enabling Hugging Face Spaces to function as a vast library of MCP servers that grant LLMs new capabilities like image editing and transcription.

255

Creating Custom Kernels for the AMD MI300

Hugging Face collaborated with AMD to develop open-source optimized kernels for the MI300X, significantly improving FP8 inference performance for Llama 3.1 405B in VLLM.

256

Reachy Mini: Open-Source Desktop Robot for AI Development

Hugging Face and Pollen Robotics have introduced Reachy Mini, an open-source, programmable desktop robot starting at $399 designed for human-robot interaction and AI experimentation.

257

Hugging Face Efficient MultiModal Data Pipeline

Hugging Face introduces a five-stage optimization process for multimodal data pipelines, utilizing a balanced knapsack packing strategy to minimize GPU idle time and padding waste.

258

SmolLM3 Release: Multilingual, Long-Context 3B Reasoner

Hugging Face has released SmolLM3, a 3B parameter model that outperforms Llama-3.2-3B and Qwen2.5-3B, featuring dual-mode reasoning and a 128k context window.

259

Hugging Face Production Infrastructure Alerting Strategies

Hugging Face utilizes specialized alerting for NAT gateway throughput, log archival success rates, and Kubernetes API health to maintain stability and cost-efficiency in its production environment.

260

NeurIPS 2025 E2LM Competition: Early Training Evaluation of Language Models

Hugging Face and partners announce the E2LM competition to develop benchmarks that can detect reasoning and scientific knowledge signals during the early stages of LLM training.

261

Training and Finetuning Sparse Embedding Models with Sentence Transformers

Hugging Face introduces a comprehensive guide for training and finetuning sparse embedding models using the Sentence Transformers library to improve hybrid search and retrieval performance.

262

NVIDIA Llama Nemotron Nano VL Release

NVIDIA has released Llama Nemotron Nano VL, an 8B Vision Language Model (VLM) optimized for high-accuracy intelligent document processing and OCR tasks.

263

Gemma 3n Release Notes: Multimodal On-Device AI

Google's Gemma 3n is now available in the open-source ecosystem, featuring native multimodality (text, image, audio, video) and memory-efficient architectures designed for local hardware execution.

264

SGLang Transformers Backend Integration

SGLang now supports Hugging Face transformers as a backend, enabling high-performance inference for any transformers-compatible model without requiring native SGLang support.

265

Fine-Tuning FLUX.1-dev with QLoRA on Consumer Hardware

Hugging Face demonstrates how to fine-tune the FLUX.1-dev model using QLoRA and the diffusers library to reduce peak VRAM usage to under 10 GB on a single consumer GPU.

266

Groq Integration with Hugging Face Inference Providers

Hugging Face has added Groq as a supported Inference Provider, allowing users to access high-speed LPU-powered inference for models like Llama 4 and QWQ-32B directly through the Hub.

267

Optimizing LLM Performance: Solving Long Prompt Blocking and Decode Slowdowns

Hugging Face explores how long prompts block request queues and slow down token generation, proposing request-parallel prefills and disaggregated prefill as solutions to reduce latency.

268

Featherless AI Integration with Hugging Face Inference Providers

Hugging Face has added Featherless AI as a supported Inference Provider, enabling serverless access to a vast catalog of open-source text and conversational models.

269

Hugging Face Kernel Hub Release

Hugging Face has introduced the Kernel Hub, a centralized repository for loading pre-compiled, optimized compute kernels directly into Python applications to accelerate GPU operations without local compilation.

270

NVIDIA Isaac GR00T N1.5: Post-Training for LeRobot SO-101 Arm

NVIDIA has released Isaac GR00T N1.5, an open foundation model for humanoid robot reasoning and skills that can be post-trained for specific robotic embodiments like the LeRobot SO-101 arm.

271

Hugging Face and NVIDIA Launch Training Cluster as a Service

Hugging Face and NVIDIA have introduced Training Cluster as a Service, a collaboration designed to provide research organizations with flexible, on-demand access to large-scale NVIDIA GPU clusters for training foundational models.

272

Hugging Face ScreenSuite Release

Hugging Face has released ScreenSuite, a comprehensive evaluation suite for GUI agents that unifies 13 benchmarks to assess Vision Language Models (VLMs) across perception, grounding, and action capabilities.

273

KV Cache Implementation in nanoVLM

Hugging Face implemented KV Caching from scratch in the nanoVLM repository, resulting in a 38% speedup in generation for their Vision Language Model.

274

Real-Time AI Sound Generation on Arm CPUs

Arm and Hugging Face demonstrate a personal sound generation tool using Stable Audio Open that enables real-time, on-device audio creation for music production workflows.

275

Holo1 and Surfer-H: Open-Source Action VLMs for GUI Automation

H Company has released Holo1, a family of Action Vision Language Models designed for precise GUI localization, and Surfer-H, a modular web agent that achieves 92.2% accuracy on real-world tasks at a cost of $0.13 per task.

276

Hugging Face TRL: Co-located vLLM for Efficient GRPO Training

Hugging Face introduces co-located vLLM in TRL, allowing training and inference to share the same GPUs to eliminate idle time and increase throughput during GRPO training.

277

SmolVLA: Efficient Vision-Language-Action Model trained on Lerobot Community Data

Hugging Face introduces SmolVLA, a compact 450M parameter open-source Vision-Language-Action model that outperforms larger models on simulation and real-world robotics tasks using community-shared data.

278

Hugging Face CodeAgents + Structure: Improving Agent Reliability via Structured Generation

Hugging Face research demonstrates that forcing CodeAgents to generate thoughts and code within a structured JSON format improves performance by 2-7 percentage points on average for capable models by eliminating parsing errors and enforcing explicit reasoning.

279

Liger GRPO and TRL Integration

A user report identifies a shape mismatch error when using Liger GRPO loss with DeepSpeed ZeRO-3 and the Qwen2.5-0.5B-Instruct model in bf16.

280

Hugging Face Tiny Agents in Python

Hugging Face has introduced Tiny Agents in Python, a lightweight agent framework powered by the Model Context Protocol (MCP) that allows LLMs to interact with external tools using minimal code.

281

Dell Enterprise Hub Update: On-Premises AI Models and Applications

Dell and Hugging Face have updated the Dell Enterprise Hub to provide a complete suite of optimized models and ready-to-deploy AI applications for Dell AI Servers and AI PCs.

282

Falcon-H1: Hybrid-Head Language Models for Efficiency and Performance

Hugging Face and TII UAE introduced Falcon-H1, a family of six open-source hybrid-head models (0.5B to 34B) that combine Transformer attention with Mamba-2 State Space Models to achieve high performance with lower memory and faster inference.

283

Falcon-Arabic: A Breakthrough in Arabic Language Models

The Technology Innovation Institute (TII) has released Falcon-Arabic, a 7B parameter model that outperforms larger Arabic LLMs in general knowledge, grammar, and reasoning across Modern Standard Arabic and regional dialects.

284

nanoVLM: A Minimalist PyTorch Toolkit for Training Vision Language Models

Hugging Face has released nanoVLM, a lightweight, pure PyTorch toolkit designed to simplify the training and understanding of Vision Language Models (VLMs) for beginners and researchers.

285

Hugging Face Diffusers Quantization Backends

Hugging Face Diffusers integrates multiple quantization backends including bitsandbytes, torchao, Quanto, GGUF, and FP8 layerwise casting to reduce the memory footprint of large diffusion models like FLUX.1-dev.

286

Microsoft and Hugging Face Expand Collaboration for Azure AI Foundry

Microsoft and Hugging Face have expanded their partnership to integrate over 10,000 open-source models into Azure AI Foundry, enabling secure, enterprise-grade deployment of diverse AI modalities.

287

Falcon-Edge: Powerful, Universal, and Fine-Tunable 1.58-bit LLMs

Hugging Face and the Falcon-LLM team have released Falcon-Edge, a series of 1.58-bit (ternary) language models in 1B and 3B parameter sizes that support both inference and fine-tuning through a new pre-training paradigm.

288

Hugging Face Transformers: Standardizing Model Definitions for Ecosystem Interoperability

Hugging Face is positioning the Transformers library as the central pivot for model definitions to ensure that any architecture supported by Transformers is automatically compatible with the broader ML ecosystem, including inference engines and training frameworks.

289

Hugging Face and Kaggle Integration for Model Access

Hugging Face and Kaggle have launched an integration that improves the discoverability and usability of Hugging Face models directly within Kaggle notebooks and model pages.

290

Hugging Face Inference Endpoints: Fast Whisper Transcriptions

Hugging Face has introduced a new OpenAI Whisper deployment option on Inference Endpoints that delivers up to 8x performance improvements in transcription speed without sacrificing accuracy.

291

Hugging Face Vision Language Models 2025 Update

Hugging Face provides a comprehensive overview of the 2024-2025 evolution of Vision Language Models (VLMs), highlighting trends in any-to-any architectures, reasoning models, and the rise of Vision-Language-Action (VLA) models for robotics.

292

LeRobot Community Datasets: Building the ImageNet of Robotics

Hugging Face is fostering a community-driven effort to create a diverse, open-source repository of robotics datasets via LeRobot to solve the generalization challenge in robotic policies.

293

Building MCP Servers with Gradio

Hugging Face has integrated the Model Context Protocol (MCP) into Gradio, allowing developers to turn Python functions into LLM-accessible tools with a single parameter change.

294

Qwen-3 Chat Template Analysis

The Qwen-3 model introduces a sophisticated chat template that enables optional reasoning, dynamic context management via rolling checkpoints, and improved tool argument serialization.

295

Intel AutoRound: Advanced Weight-Only Quantization for LLMs and VLMs

Intel has introduced AutoRound, a weight-only post-training quantization method that uses signed gradient descent to enable high-accuracy low-bit quantization (INT2-INT8) for LLMs and VLMs.

296

Llama Guard 4 and Llama Prompt Guard 2 Release

Meta has released Llama Guard 4, a 12B multimodal safety model, and Llama Prompt Guard 2, a series of classifiers for detecting prompt injections and jailbreaks.

297

PipelineRL: Optimizing LLM Reinforcement Learning via Inflight Weight Updates

Hugging Face and ServiceNow Research have open-sourced PipelineRL, an experimental RL implementation that uses inflight weight updates to eliminate the trade-off between inference throughput and on-policy data collection.

298

Hugging Face Tiny Agents: Building MCP-Powered Agents in 50 Lines of Code

Hugging Face demonstrates how to build a functional AI agent using the Model Context Protocol (MCP) and the InferenceClient, reducing the core agent logic to a simple while loop.

299

Finetuning olmOCR for Faithful Document Extraction

TNG has released a fine-tuned version of olmOCR-7B-0225-preview that preserves headers and footers, making it suitable for business applications like invoice parsing.

300

Optimizing LLM Performance: Prefill and Decode for Concurrent Requests

Hugging Face (via TNG) explains how managing the prefill and decode phases of token generation through strategies like continuous batching and chunked prefill can optimize GPU utilization and increase token throughput by up to 50%.