Releases

Hugging Face released Candle, a minimalist Rust framework for fast and portable machine learning.

Meta released AudioCraft with MusicGen, AudioGen and EnCodec components for generative audio research.

CoreWeave released Tensorizer, an open-source tensor serialization tool that streams models from S3, HTTP, and file storage to reduce PyTorch model startup time.

Microsoft announces a common plugin model spanning ChatGPT, Bing, Dynamics 365 Copilot and Microsoft 365 Copilot.

Anthropic expanded Claude's API context window from 9K to 100K tokens.

Anthropic made Claude available through an API and chat interface in its developer console.

OpenAI made the Whisper speech-to-text model available through its API.

OpenAI released the Chat Completions API with gpt-3.5-turbo for developers.

Microsoft releases Semantic Kernel to help developers integrate language models, plugins, memory and planning into applications.

Cohere introduced a conversational Chat API that evolved to support RAG, citations, tools and structured outputs.

Hugging Face released Parameter-Efficient Fine-Tuning tools for adapting large models with methods such as LoRA.

Cohere made relevance reranking available through a dedicated API endpoint requiring only a small search-stack integration.

Cohere introduced a multilingual cross-encoder Rerank endpoint for improving keyword and semantic search results.

Meta open-sourced AITemplate, a Python framework that transforms neural networks into optimized GPU code.

Hugging Face released Text Generation Inference, a production server optimized for large-language-model generation.

Hugging Face released Transformer Reinforcement Learning tooling for post-training language models with supervised, preference, and reinforcement-learning methods.

Hugging Face launched Diffusers, an open-source library providing modular pipelines for diffusion-based generative models.

Hugging Face launched Evaluate as a unified framework for model and dataset evaluation.

Microsoft releases the Responsible AI Dashboard, combining fairness, interpretability, error analysis and causal tools in one interface.

Baidu released an inference and deployment toolkit for AI models.

NVIDIA expanded TAO Toolkit integration for fine-tuning and optimizing pretrained production AI models.

NVIDIA made its end-to-end GPU-accelerated recommender AI framework generally available.

NVIDIA made its multilingual speech AI SDK generally available.

Hugging Face introduced SafeTensors as a fast tensor serialization format designed to avoid arbitrary-code execution.

Meta open-sourced TorchRec, a PyTorch library for large-scale recommendation systems.

OpenAI released a new endpoint and model families for text and code embeddings.

Oracle released managed time-series forecasting for business and operational signals.

Oracle released managed multivariate time-series anomaly detection.

Oracle released managed image analysis, object detection and custom vision models.

Oracle released a managed natural-language processing API for classification, entity extraction and sentiment.

Oracle released managed automatic speech recognition and transcription.

Cohere's hosted API made semantic text representations available to enterprise developers.

Cohere's initial commercial platform included text classification for enterprise NLP workflows.

Cohere made its hosted language-model platform available to businesses alongside its Series A financing.

Hugging Face launched Optimum to optimize Transformers models for specialized hardware and inference runtimes.

Hugging Face launched Accelerate to let PyTorch code run across distributed hardware with minimal changes.

Transformers v4 began its release cycle with broad framework and architecture support and a standardized model API.

Hugging Face introduced a hosted API for running models from the Hub without managing serving infrastructure.

OpenAI launched a general-purpose API for accessing its language models.

Hugging Face released a unified open-source library and catalog for loading, processing, streaming, and sharing datasets.

Hugging Face introduced its high-performance Rust-backed Tokenizers library for production-scale text preprocessing.

Microsoft open-sources DeepSpeed with the ZeRO optimizer to reduce the compute and memory cost of large-model training.

Baidu released its open NLP development library for PaddlePaddle.

Facebook AI released Detectron2, a modular object-detection and segmentation library.

Hugging Face renamed pytorch-transformers to Transformers and expanded the library’s framework and architecture support.

Baidu released an open OCR toolkit built on PaddlePaddle.

Facebook AI released AI Habitat, a simulation platform for embodied AI research.

Google releases MediaPipe as a cross-platform framework for building perception and multimodal ML pipelines.

Microsoft open-sources InterpretML for training interpretable models and explaining black-box predictions.

AWS released optimized Docker images for deep-learning training and inference.

Gradio launched as an open-source Python framework for rapidly building browser interfaces around machine-learning models.

Microsoft open-sources ONNX Runtime as a cross-platform high-performance inference engine.

Google Research releases JAX for composable automatic differentiation and accelerator-backed numerical computing.

Hugging Face released an open-source PyTorch implementation and pretrained checkpoints for BERT, the project that evolved into the Transformers library.

NVIDIA launched RAPIDS as an open-source GPU-accelerated data science and machine-learning suite.

Microsoft releases the first public preview of ML.NET, bringing cross-platform machine learning to .NET developers.

Microsoft researchers introduce Fairlearn to assess and mitigate unfairness in machine-learning systems.

OpenAI released code for training reinforcement-learning agents from human feedback.

OpenAI released Baselines, open-source high-quality implementations of reinforcement-learning algorithms.

Facebook released the Caffe2 deep-learning framework for scalable cross-platform training and deployment.

Facebook AI Research released FAISS for efficient similarity search and clustering of dense vectors.

Facebook AI Research released PyTorch, a Python-first tensor and dynamic neural-network framework.

Facebook AI Research released fastText, a library for efficient text classification and word representations.

Scale launched an API that let developers submit human-powered data tasks with one line of code, establishing the foundation of its AI data business.

Amazon open-sourced its Deep Scalable Sparse Tensor Network Engine for recommendation workloads.

Microsoft releases its Computational Network Toolkit under an open-source license for deep neural-network training.

Amazon opened Alexa to developers through the Alexa Skills Kit and Alexa Voice Service.

NVIDIA released the first version of its GPU-accelerated deep neural network primitives library.

NVIDIA introduced CUDA and a C programming environment, opening its GPUs to general-purpose parallel computing.