Cohere and Hugging Face introduce Aya Expanse, a multilingual language model covering 119 languages. The model improves performance on translation, text generation, and reasoning tasks across underrepresented languages, with public benchmarks released.
OpenAI simplifies and stabilizes continuous-time consistency models, achieving sample quality comparable to leading diffusion models while requiring only two sampling steps.
Hugging Face introduces SynthID Text, a technique to identify and trace AI-generated text. The system embeds imperceptible markers in generated content, enabling detection of origin without degrading quality.
Hugging Face releases Outlines-core 0.1.0, a library for structured generation in Rust and Python. The tool constrains language model outputs to specific formats (JSON, regex, grammars). Stable version with multi-language support.
Transformers.js v3 adds WebGPU support for browser-side GPU acceleration, integrates new models and tasks, and improves inference performance. The release includes optimizations for running AI models locally in the browser.
Hugging Face integrates Stable Diffusion 3.5 Large into its Diffusers library. The image generation model is now available through the official API with full optimized pipeline support.
OpenAI introduces MLE-bench, a benchmark for evaluating AI agents' performance on machine learning engineering tasks. The tool measures agents' ability to complete complex ML work.
Hugging Face released a comprehensive security audit of Gradio 5, identifying and fixing critical vulnerabilities before launch. The review covers authentication, file handling, code injection, and dependencies.
Gradio 5 launches with major UI redesign, improved components, and better AI model integration. The platform simplifies ML app deployment and strengthens sharing capabilities.
Hugging Face introduces dynamic speculation for assisted generation, accelerating text generation through adaptive token speculation. The method adjusts speculated tokens based on model confidence, reducing latency without quality loss.
OpenAI launches Canvas, a new interface for writing and coding with ChatGPT. Canvas provides a dedicated space to view and edit generated content in real-time, separating conversation from final output. Available to ChatGPT Plus and Teams users.
OpenAI secures $6.6 billion in new funding and establishes a $4 billion credit facility with JPMorgan Chase, Citi, Goldman Sachs, Morgan Stanley, Santander, Wells Fargo, SMBC, UBS, and HSBC.
OpenAI introduces prompt caching in its API, automatically reducing costs for inputs the model has recently processed. This feature decreases latency and fees for repetitive requests.
OpenAI introduces model distillation through its API: fine-tune a cost-efficient model using outputs from a large frontier model, all on the OpenAI platform.
OpenAI launches a new multimodal moderation model built on GPT-4o with improved accuracy for detecting harmful text and images. Developers can now build more robust moderation systems.
Hugging Face introduces extreme quantization method reducing LLMs to 1.58bit. The technique simplifies fine-tuning of quantized models, enabling efficient training on limited hardware.
Hugging Face releases Accelerate 1.0.0, a library for training and inference on multi-GPU/TPU setups. Stable release with frozen API, PyTorch/JAX support, and DeepSpeed integration.
Hugging Face improves training efficiency by combining packing (grouping short sequences) with Flash Attention 2. This technique reduces unnecessary padding and accelerates attention computation, increasing training throughput without degrading model quality.
Hugging Face and Google Cloud announce Llama 3.1 405B deployment on Vertex AI. The 405-billion-parameter model is available via Vertex AI API with fine-tuning and RAG support.
OpenAI releases SWE-bench Verified, a human-validated subset of SWE-bench to more reliably evaluate AI models' ability to solve real-world software issues.
Hugging Face introduces Falcon Mamba, a 7B attention-free model based on Mamba architecture. It matches standard attention-based models in performance while delivering faster inference and reduced memory consumption.
Zico Kolter joins OpenAI's Board of Directors. He also joins the Safety & Security Committee, strengthening governance expertise in AI safety and alignment.
XetHub, a data management platform for ML, joins Hugging Face. The acquisition strengthens versioning and collaboration capabilities for datasets and models within the Hugging Face ecosystem.
OpenAI releases GPT-4o System Card, documenting the model's capabilities, limitations, and safety measures. The document details benchmark performance, identified risks, and deployed mitigations.
Hugging Face and NVIDIA launch serverless inference integration via NVIDIA NIM. Users can deploy models without managing infrastructure, with automatic scaling and NVIDIA optimizations.
Hugging Face releases TGI Multi-LoRA, enabling a single deployment to serve 30 models through LoRA adapters. Reduces infrastructure costs and simplifies multi-model management in production.
OpenAI releases compliance and administrative tools for ChatGPT Enterprise: Compliance API integrations, SCIM support, and GPT controls to manage user access, data security, and compliance programs at scale.
Hugging Face introduces SmolLM, a family of lightweight language models designed for speed and efficiency. The models deliver strong performance despite their reduced size, optimized for fast inference and low energy consumption.
NuminaMath won the 1st AIMO Progress Prize by combining fine-tuning and data synthesis techniques on base models. The solution leverages advanced mathematical reasoning to solve complex olympiad-level problems.
Google Cloud makes TPUs available to Hugging Face users through direct integration. Users can access TPU accelerators for model training and inference without complex setup.
Google releases Gemma 2, an open LLM available in 9B and 27B variants. The model delivers performance comparable to proprietary models in its class, emphasizing efficiency and accessibility for the open-source community.
OpenAI acquires Rockset, a vector database and real-time indexing platform. The acquisition strengthens OpenAI's infrastructure capabilities to support large-scale AI applications.
OpenAI introduces Consistency Models, an approach that accelerates image, audio, and video generation by reducing the number of iterations required by traditional diffusion models.
Hugging Face introduces BigCodeBench, a next-generation benchmark for evaluating code generation models. It supersedes HumanEval with expanded coverage and improved metrics to measure code generation capabilities.
OpenAI appoints retired U.S. Army General Paul M. Nakasone to its Board of Directors. He will join the Safety and Security Committee, bringing cybersecurity expertise from his tenure as NSA and Cyber Command leader.
Hugging Face integrates Stable Diffusion 3 into its Diffusers library. The image generation model is now available through the open-source infrastructure with full pipeline support and performance optimizations.
OpenAI identified 16 million patterns in GPT-4's computations using scaled sparse autoencoder techniques. This breakthrough enables extraction and understanding of the model's internal concepts.
Hugging Face releases a benchmark for evaluating text generation inference system performance. The study measures latency, throughput, and resource utilization across different model architectures and configurations.
Hugging Face releases Falcon 2, an 11B parameter model trained on 5000B+ tokens across 11 languages. Available as both text and multimodal vision language model (VLM).
Meta releases CyberSecEval 2, a comprehensive evaluation framework measuring cybersecurity risks and capabilities of LLMs. The tool tests malicious code generation, vulnerability exploitation, and attack defense across models including Llama.