✷ 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.
Stable Diffusion 3.5 Large Integration with Diffusers
Hugging Face has integrated Stable Diffusion 3.5 Large, an 8B parameter model available in standard and timestep-distilled versions, into the Diffusers library.
Hugging Face partners with Protect AI to add Guardian scanner for model security
Hugging Face partnered with Protect AI to embed the Guardian scanner into the Hub, automatically detecting dangerous model serialization exploits and improving security for the entire ML community.
Transformers.js v3 release adds WebGPU acceleration, expanded model support, and server‑side JavaScript compatibility
Transformers.js v3 adds WebGPU acceleration, new quantization formats, support for 120 model architectures, and Node.js/Deno/Bun compatibility, enabling fast, on‑device inference in browsers and JavaScript runtimes.
Llama 3.2 in Keras
Llama 3.2 is fully supported in Keras via keras-hub, allowing users to load Hugging Face checkpoints and run models across JAX, PyTorch, or TensorFlow backends.
Hugging Face Transformers Gradient Accumulation Fix
Hugging Face has updated the Transformers Trainer to ensure gradient accumulation is mathematically equivalent to full batch training by correcting how losses are averaged across batches.
Gradio 5 Security Review
Hugging Face conducted a comprehensive security audit of Gradio 5 with Trail of Bits, fixing all identified vulnerabilities to ensure machine learning applications are safe by default.
AMD EPYC Turin CPU delivers 2× LLM inference throughput over Genoa
AMD’s 5th‑gen EPYC Turin CPU delivers roughly double the LLM inference throughput of Genoa, enabling lower latency and higher throughput for Hugging Face workloads.
Scaling AI Data Processing with Hugging Face and Dask
Hugging Face and Dask enable the scaling of AI-based data processing from small local samples to hundreds of millions of rows using distributed computing and multi-GPU parallel inference.
Gradio 5 Release Notes
Hugging Face has released Gradio 5, a production-ready framework for building performant, scalable, and secure machine learning web applications using Python.
Hugging Face Transformers 4.45.0 Dynamic Speculative Decoding
Hugging Face and Intel Labs introduced dynamic speculative decoding in Transformers 4.45.0, accelerating text generation by up to 2.7x by dynamically adjusting the number of draft tokens based on model confidence.
Improving Parquet Deduplication on Hugging Face Hub
Hugging Face is optimizing its storage architecture to improve Parquet file deduplication, proposing content-defined row groups to reduce storage overhead during dataset updates.
Open FinLLM Leaderboard launch – comprehensive zero‑shot benchmark for financial language models
Hugging Face launched the Open FinLLM Leaderboard, a zero‑shot benchmark covering 40 finance‑specific tasks across seven categories to evaluate LLM readiness for real‑world financial applications.
Chinese AI Global Expansion Analysis
Chinese AI companies are accelerating international expansion due to domestic market saturation, intense price wars, and regulatory pressures, targeting Southeast Asia, the Middle East, and Western consumer markets.
BenCzechMark: A Comprehensive Evaluation Suite for Czech LLMs
Hugging Face and academic partners have released BenCzechMark, the first comprehensive evaluation suite for Czech language models, featuring 50 tasks across 9 categories and a novel duel-based scoring mechanism.
Converting Vertex-Colored Meshes to Textured Meshes
Hugging Face introduces a method and the InstantTexture library to convert vertex-colored 3D meshes into UV-mapped, textured meshes for better application compatibility.
Llama 3.2 Release Notes: Multimodal Vision and On-Device Small Language Models
Meta has released Llama 3.2, introducing multimodal vision capabilities in 11B and 90B sizes and lightweight 1B and 3B text-only models optimized for on-device deployment.
Hugging Face Daily Papers Features Guide
Hugging Face's Daily Papers page provides a community-curated hub for AI research, featuring tools for author claiming, paper submission, and direct interaction between researchers and developers.
FineVideo Dataset Release
Hugging Face has released FineVideo, a high-quality open video dataset containing 43k videos (3.4k hours) with rich, structured annotations for video understanding and generative AI training.
Optimizing and Deploying Hugging Face Models with Optimum-Intel and OpenVINO GenAI
Hugging Face and Intel provide a streamlined workflow using Optimum-Intel and OpenVINO GenAI to optimize and deploy Transformers models on Intel hardware, specifically targeting edge and client-side C++ and Python environments.
Fine-tuning LLMs to 1.58-bit: Extreme Quantization with BitNet
Hugging Face demonstrates that existing LLMs, such as Llama 3 8B, can be fine-tuned to 1.58-bit ternary precision using a dynamic warmup quantization strategy, significantly reducing memory and energy requirements while maintaining strong performance.
Hugging Face SQL Console for Datasets
Hugging Face has introduced a browser-based SQL Console powered by DuckDB WASM that allows users to query, filter, and transform datasets directly on the Hub without backend dependencies.
HuggingChat Community Tools Release
Hugging Face has introduced Community Tools on HuggingChat, allowing users to integrate any public Hugging Face Space as a tool for LLMs to use directly within the chat interface.
Accelerate 1.0.0 Release Candidate Announcement
Accelerate 1.0.0 release candidates add FP8, DeepSpeed multi‑model, torch.compile, and new data‑loader/pipeline features while stabilizing the API for large‑scale training and inference.
Hugging Face and TruffleHog Partnership for Secret Scanning
Hugging Face has partnered with Truffle Security to integrate TruffleHog's secret scanning capabilities into its automated pipeline and provide a native scanner for users to proactively scan their own account data.
LeRobotDataset video‑encoding format reduces robotics dataset size and speeds up training
Hugging Face released the LeRobotDataset video‑encoding format, shrinking robotics visual data to about 14 % of its original size while keeping loading speed and training performance intact.
Hugging Face blog post highlights five under‑rated Hub tools and a free semantic‑search use case
Hugging Face announced five under‑rated Hub tools—ZeroGPU, multi‑process Docker, Gradio API, webhooks, and Nomic Atlas—and showed how to combine them into a free, auto‑updating semantic‑search app for Reddit data.
Hugging Face Training Efficiency: Packing with Flash Attention 2
Hugging Face has introduced boundary-aware packing for instruction tuning examples, enabling up to 2x training throughput increase and 20% peak memory reduction when used with Flash Attention 2.
Deploying Meta Llama 3.1 405B on Google Cloud Vertex AI
Hugging Face provides a guide for programmatically deploying the FP8 quantized version of Meta Llama 3.1 405B on Google Cloud Vertex AI using Text Generation Inference (TGI) and A3 machine series.
Hugging Face Infini-Attention Reproduction Analysis
Hugging Face's attempt to reproduce Infini-Attention found that while gating convergence can be improved, the method's performance degrades with increased memory compression and remains less reliable than Ring Attention, YaRN, or RoPE scaling.
Introduction to ggml
ggml is a lightweight, C/C++ machine learning library optimized for Transformer inference and on-device LLM execution across diverse hardware backends.
Hugging Face Unified Tool Use API
Hugging Face has introduced a unified tool use API that allows developers to use the same code to implement tool calling across Mistral, Cohere, NousResearch, and Llama models.
Falcon Mamba 7B Release Notes
The Technology Innovation Institute (TII) has released Falcon Mamba 7B, the first large-scale pure State Space Language Model (SSLM) that matches the performance of state-of-the-art transformer models while eliminating attention-based memory scaling issues.
Hugging Face acquires XetHub to upgrade Hub storage and collaboration
Hugging Face acquired XetHub to replace Git LFS with a more efficient storage backend, enabling incremental updates, trillion‑parameter model support, and better collaboration on massive AI datasets.
Hugging Face TextImage Augmentation pipeline release
Hugging Face and Albumentations AI released a TextImage Augmentation pipeline that jointly modifies document images and their text, enabling realistic synthetic data generation and robust fine‑tuning of vision‑language models on limited document datasets.
Hugging Face 2024 Security Feature Highlights
Hugging Face has detailed its 2024 security landscape, introducing a suite of default protections for all users and advanced governance controls for Enterprise Hub users.
Google releases Gemma 2 2B, ShieldGemma, and Gemma Scope
Google released Gemma 2 2B, ShieldGemma safety classifiers, and Gemma Scope sparse autoencoders, expanding open‑source LLM capabilities, moderation tools, and interpretability resources.
Quanto quantization cuts memory for Transformer diffusion pipelines
Hugging Face quantization (Quanto) reduces GPU memory for Transformer diffusion models from ~12 GB to ~5 GB with minimal latency and quality impact.
Hugging Face NVIDIA NIM API (serverless) launch and deprecation
Hugging Face launched the NVIDIA NIM API (serverless) for Enterprise Hub users, enabling pay‑as‑you‑go, serverless inference of open‑source LLMs on NVIDIA DGX Cloud H100 GPUs via an OpenAI‑compatible API.
LAVE: Zero-shot VQA Evaluation on Docmatix with LLMs
Hugging Face introduces LAVE (LLM-Assisted VQA Evaluation) to address the rigidity of traditional VQA metrics, demonstrating a 50% accuracy gain in evaluating zero-shot performance on the Docmatix dataset.
Llama 3.1 Release Notes: Multilinguality, Long Context, and 405B Model
Meta has released Llama 3.1, featuring models in 8B, 70B, and 405B sizes with 128K context length, multilingual support for 8 languages, and a permissive license allowing synthetic data generation.
Running Mistral 7B with Core ML
Hugging Face demonstrates how to run Mistral 7B on Mac using new Core ML features from WWDC 24, achieving a model size reduction to under 4GB using 4-bit block-wise quantization.
Docmatix Dataset Release
Hugging Face has released Docmatix, a Document Visual Question Answering (DocVQA) dataset featuring 2.4 million images and 9.5 million Q/A pairs, providing a 240x increase in scale over previous datasets.
TGI Multi-LoRA: Deploy Once, Serve 30 Models
Hugging Face introduces Multi-LoRA serving in Text Generation Inference (TGI), allowing organizations to deploy a single base model and dynamically serve dozens of specialized fine-tuned adapters to reduce cost and operational complexity.
Argilla SDK Chatbot with distilabel – End‑to‑End Tutorial
Hugging Face released a tutorial showing how to build an Argilla 2.0 chatbot using distilabel‑generated synthetic data, fine‑tuned embeddings, lancedb vector storage, and a Gradio app deployed on Spaces.
SmolLM Release: High-Performance Small Language Models
Hugging Face introduces SmolLM, a family of state-of-the-art small language models (135M, 360M, and 1.7B parameters) trained on a meticulously curated high-quality dataset.
NuminaMath 7B TIR wins AIMO Progress Prize – technical recap
NuminaMath 7B TIR won the first AIMO Progress Prize by solving 29 of 50 hidden math problems, showcasing the power of a two-stage fine‑tuning recipe, large high‑quality math data, and a self‑consistency with tool‑integrated reasoning inference strategy.
Hugging Face PII Detection Experiment with Presidio
Hugging Face is experimenting with integrating Microsoft Presidio into the Dataset Hub to provide automatic PII detection reports, helping practitioners identify and mitigate privacy risks in ML datasets.
TRL adds Direct Preference Optimization support for Vision‑Language Models
Hugging Face added Direct Preference Optimization (DPO) support for Vision‑Language Models in the TRL library, enabling fine‑tuning of models like Idefics‑2 with preference data using bfloat16 quantization and LoRA to fit on a single GPU.
Hugging Face and KerasHub Integration
Hugging Face and KerasHub now share a model save format, allowing KerasHub users to directly load over 300,000 Transformers library models from the Hugging Face Hub.
Google Cloud TPUs on Hugging Face Inference Endpoints and Spaces
Hugging Face has integrated Google Cloud TPU v5e support into Inference Endpoints and Spaces, enabling users to deploy and scale AI models with cost-effective, high-performance hardware.