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.

351

Open-R1: A Fully Open Reproduction of DeepSeek-R1

Hugging Face has launched the Open-R1 project to systematically reconstruct the data and training pipeline of DeepSeek-R1, aiming to provide the open-source community with the missing datasets and code for reasoning models.

352

Hugging Face Inference Providers Integration

Hugging Face has integrated four serverless inference providers—fal, Replicate, SambaNova, and Together AI—directly into the Hub to provide unified, model-centric serverless inference.

353

State of Open Video Generation Models in Diffusers

Hugging Face provides a comprehensive overview of open video generation models and introduces a suite of Diffusers optimizations that can reduce VRAM requirements for models like HunyuanVideo from 60GB to approximately 6.5GB.

354

smolagents Vision Support Update

Hugging Face has added native vision support to smolagents, enabling the use of Vision Language Models (VLMs) in agentic pipelines for tasks like autonomous web browsing.

355

NVIDIA KVPress: Toolkit for KV Cache Compression in Long-Context LLMs

NVIDIA has released KVPress, a Python toolkit that implements state-of-the-art KV cache compression techniques to reduce the memory footprint and increase decoding speed of long-context Large Language Models.

356

SmolVLM 256M and 500M Release Notes

Hugging Face introduces SmolVLM-256M and SmolVLM-500M, delivering highly efficient Vision Language Models that maintain strong multimodal performance in a significantly reduced parameter footprint.

357

Hugging Face and FriendliAI Partnership for Model Deployment

Hugging Face has integrated FriendliAI's inference infrastructure into the Hugging Face Hub, enabling one-click deployment of generative AI models to high-performance endpoints.

358

Hugging Face Organization Blog Articles Feature

Hugging Face now allows organizations subscribed to Enterprise Hub to publish blog articles directly to their organization profiles.

359

Hugging Face Text Generation Inference (TGI) Multi-Backend Support

Hugging Face has introduced a multi-backend architecture for Text Generation Inference (TGI), allowing users to utilize a unified frontend to deploy LLMs via various execution engines like TensorRT-LLM and vLLM.

360

Hugging Face Transformers timm Integration

Hugging Face has introduced the TimmWrapper, allowing any model from the PyTorch Image Models (timm) library to be used seamlessly within the transformers ecosystem for inference, quantization, and fine-tuning.

361

Train 400x faster Static Embedding Models with Sentence Transformers

Hugging Face introduces a method to train static embedding models that run 100x–400x faster on CPU while retaining at least 85% of the quality of models like all‑mpnet‑base‑v2, releasing two models (static‑retrieval‑mrl‑en‑v1 and static‑similarity‑mrl‑multilingual‑v1) with training scripts and evaluation results.

362

Hugging Face AI Agents Ethics and Framework Analysis

Hugging Face provides a comprehensive framework for understanding AI agents, arguing that risks increase with autonomy and recommending against the development of fully autonomous agents.

363

Hugging Face releases vdr-2b-multi-v1 multilingual visual document retrieval model

Hugging Face has introduced vdr-2b-multi-v1, a multilingual embedding model for visual document retrieval that enables searching complex documents without OCR by encoding page screenshots into dense vectors.

364

Hugging Face Open LLM Leaderboard: CO₂ Emissions and Model Performance Analysis

Hugging Face reveals that community fine-tuned models often exhibit higher carbon efficiency than official releases due to increased conciseness and improved instruction following.

365

Hugging Face smolagents Release

Hugging Face has launched smolagents, a lightweight library that enables LLMs to perform complex tasks by writing actions as executable code rather than JSON, improving composability and generality.

366

Visualize and understand GPU memory in PyTorch – Hugging Face Blog Summary

Hugging Face’s blog post explains how to visualize GPU memory usage in PyTorch using torch.cuda.memory tools, breaks down memory components during model training, and provides formulas to estimate total memory requirements.

367

NVIDIA LogitsProcessorZoo: Controlling Language Model Generation with Modular Logits Processors

NVIDIA's LogitsProcessorZoo provides modular logits processors for Hugging Face Transformers that let developers control generation length, enforce phrase inclusion, cite prompt content, and restrict outputs to multiple-choice choices.

368

Big Bench Audio Release

Artificial Analysis has released Big Bench Audio, a dataset of 1,000 audio questions designed to evaluate the reasoning capabilities of audio language models, revealing a significant performance gap between text and speech reasoning.

369

ModernBERT Release Notes

Hugging Face, Answer.AI, and LightOn have released ModernBERT, a state-of-the-art encoder-only model family that improves upon BERT's speed, accuracy, and context length to 8,192 tokens.

370

Bamba-9B: Inference-Efficient Hybrid Mamba2 Model

Bamba-9B, a hybrid Mamba2 model from IBM, Princeton, CMU, and UIUC trained on 2.2T open tokens, achieves 2.5x throughput and 2x latency improvements over Llama 3.1 8B in vLLM and is immediately usable in transformers, vLLM, TRL, and llama.cpp.

371

Benchmarking Language Model Performance on 5th Gen Xeon at GCP

Hugging Face benchmarked text embedding and generation on Google Cloud's C4 (5th‑gen Xeon) and N2 (3rd‑gen Xeon) instances, finding C4 delivers 10‑24× higher embedding throughput and 2.3‑3.6× higher generation throughput, yielding 7‑19× and 1.7‑2.9× total‑cost‑of‑ownership advantages respectively.

372

Falcon 3 release notes / what's new

Technology Innovation Institute (TII) has released Falcon 3, a family of decoder-only large language models under 10 billion parameters designed for high efficiency and enhanced science, math, and coding capabilities.

373

Hugging Face Synthetic Data Generator

Hugging Face has introduced the Synthetic Data Generator, a no-code application that allows users to create custom text classification and chat datasets using natural language prompts.

374

LeMaterial v1.0: LeMat-Bulk dataset release

LeMaterial v1.0 launches as an open-source initiative releasing the LeMat-Bulk dataset, which unifies 6.7M entries from Materials Project, Alexandria, and OQMD into a standardized format with seven properties to accelerate materials discovery.

375

Hugging Face Open Preference Dataset for Text-to-Image Generation

The Data is Better Together community has released an Apache 2.0 licensed open preference dataset for text-to-image generation to address the lack of open-source preference data for model alignment.

376

Hugging Face Models in Amazon Bedrock Marketplace

Hugging Face has integrated 83 open models into the Amazon Bedrock Marketplace, allowing AWS customers to deploy open models on managed infrastructure while maintaining compatibility with Bedrock APIs.

377

PaliGemma 2 Release Notes

Google has released PaliGemma 2, a new family of vision language models that combine the SigLIP image encoder with the Gemma 2 text decoder across three parameter sizes and multiple input resolutions.

378

How good are LLMs at fixing their mistakes? A chatbot arena experiment with Keras and TPUs

Hugging Face tested several sub‑10B LLMs on a simple calendar‑API task and found that Gemma 2 9B consistently fixed mistakes with minimal prompting, while smaller and older models struggled or required many corrective turns.

379

AraGen Benchmark and Leaderboard: Introducing 3C3H Evaluation for Arabic LLMs

Hugging Face introduced AraGen, a dynamic benchmark and leaderboard for Arabic LLMs that uses the 3C3H measure to evaluate correctness, completeness, conciseness, helpfulness, honesty, and harmlessness.

380

Hugging Face CFM Case Study: Fine-tuning Small Models with LLM Insights

Capital Fund Management (CFM) improved financial Named Entity Recognition (NER) accuracy by up to 6.4% and reduced inference costs by up to 80x by using Llama 3.1 to assist in labeling data for fine-tuning compact models like GLiNER and SpanMarker.

381

Open Source Developers Guide to the EU AI Act

The Hugging Face guide explains how the EU AI Act applies to open source AI developers, outlining obligations for limited‑risk AI systems and non‑systemic‑risk general purpose AI models and pointing to tools for compliance.

382

Hugging Face Hub Storage Rearchitecture

Hugging Face is redesigning its upload and download architecture by introducing a content-addressed store (CAS) to enable byte-level deduplication and improve global transfer speeds for massive AI models.

383

SmolVLM release notes / what's new

Hugging Face introduces SmolVLM, a 2B parameter Vision Language Model (VLM) that is fully open-source and optimized for low memory footprints and high throughput on edge devices.

384

You could have designed state of the art positional encoding

The Hugging Face blog post walks through an iterative design of positional encoding for transformers, showing how sinusoidal encoding leads to Rotary Positional Encoding (RoPE) and why it matters for modeling token relationships.

385

Hugging Face and LLM-jp Launch Open Japanese LLM Leaderboard

Hugging Face and LLM-jp have introduced the Open Japanese LLM Leaderboard, a transparent evaluation platform featuring over 20 datasets to benchmark the performance of Japanese large language models.

386

Hugging Face Introduces Content-Defined Chunking to Improve Storage Efficiency

Hugging Face announced a content-defined chunking storage approach via its Xet team that reduces storage and transfer costs for large model and dataset files by only uploading modified chunks.

387

Faster Text Generation with Self-Speculative Decoding

Hugging Face introduces self-speculative decoding via LayerSkip, a method that uses a single LLM's early layers for drafting and later layers for verification to increase generation speed and reduce memory overhead.

388

FlagEval Debate: A New Multilingual LLM Evaluation Framework

BAAI has launched FlagEval Debate, a dynamic evaluation platform where LLMs compete in multilingual debates to better assess reasoning, logic, and adversarial capabilities.

389

Hugging Face Judge Arena: Benchmarking LLMs as Evaluators

Hugging Face has launched Judge Arena, a crowdsourced platform that uses human voting to determine which LLMs are most effective as evaluators for grading other AI-generated responses.

390

Hugging Face Hub Dataset Sharing for Researchers

Hugging Face Hub provides a comprehensive platform for hosting and sharing large-scale ML datasets with integrated tools for exploration, security, and community engagement.

391

Hugging Face PyCharm Integration

Hugging Face has integrated its Hub directly into PyCharm Professional, allowing developers to discover, insert, and manage machine learning models without leaving their IDE.

392

Argilla 2.4 release notes / what's new

Argilla 2.4 introduces a no-code UI for importing Hugging Face Hub datasets to build fine-tuning and evaluation datasets through human feedback.

393

Universal Assisted Generation: Faster Decoding with Any Assistant Model

Hugging Face and Intel Labs introduced Universal Assisted Generation (UAG), a method that accelerates LLM inference by 1.5x-2.0x by allowing any small model to act as an assistant regardless of its tokenizer.

394

Digital Green Farmer.chat: Bolstering RAG with LLM-as-a-Judge

Digital Green implemented an LLM-as-a-judge evaluation framework for Farmer.chat, a RAG-based agricultural chatbot, to objectively measure RAG accuracy and optimize model selection across 340k queries.

395

Aya Expanse Release: Advancing Multilingual LLM Performance

Hugging Face and Cohere For AI have released Aya Expanse, a family of 8B and 32B open-weight models that set new state-of-the-art benchmarks for multilingual performance.

396

HUGS launch: zero‑configuration, hardware‑optimized inference for open LLMs

Hugging Face launched HUGS, a zero‑configuration, hardware‑optimized inference service for open‑source LLMs that runs on NVIDIA, AMD, and soon AWS Inferentia and Google TPUs, enabling enterprises to host models in‑house with an OpenAI‑compatible API.

397

CinePile 2.0 release: adversarial refinement boosts video QA dataset quality

CinePile 2.0 introduces an adversarial refinement pipeline that upgrades weak QA pairs into vision‑dependent questions, releasing both the improved dataset and the full code, and shows significant performance gains for commercial and open‑source video‑LLMs.

398

SynthID Text Integration in Transformers v4.46.0

Google DeepMind and Hugging Face have integrated SynthID Text into Transformers v4.46.0, providing a method to apply imperceptible watermarks to AI-generated text for detection via trained classifiers.

399

Deploying Speech-to-Speech on Hugging Face Inference Endpoints

Hugging Face provides a guide for deploying its Speech-to-Speech (S2S) pipeline using custom Docker images on Inference Endpoints to handle high computational demands and reduce latency.

400

Outlines-core 0.1.0 release notes / what's new

Hugging Face and dottxt have released outlines-core 0.1.0, a Rust port of the Outlines core algorithms for structured generation that improves index compilation speed and portability.