Releases

Mistral AI released Mistral 7B, its first open-weight language model, under the Apache 2.0 license.

Meta introduced Emu as the image-generation technology powering new generative AI experiences.

Microsoft Research releases Phi-1.5, a compact language model trained primarily on synthetic textbook-quality data.

Meta released code-specialized Llama models for generation, completion and instruction following.

Hugging Face released IDEFICS, an 80B open-access visual-language model supporting sequences of images and text.

Meta released SeamlessM4T for multilingual speech-to-speech, speech-to-text, text-to-speech and text translation.

xAI completed Grok-0, a 33-billion-parameter research prototype that preceded the first public Grok model.

Anthropic released Claude Instant 1.2 through its API with gains in reasoning, coding, multilingual performance, and safety.

Meta released Llama 2 models in 7B, 13B and 70B sizes for research and commercial use.

Anthropic released Claude 2 with improved reasoning, coding, safety, and long-context capabilities.

MosaicML released the MPT-30B base, instruct and chat models with commercial-use terms.

Meta released MMS speech-recognition and synthesis models covering more than 1,100 languages.

Google launches PaLM 2, improving multilingual, reasoning and coding capabilities across Google products and Cloud.

Meta released ImageBind, a joint embedding model spanning images, text, audio, depth, thermal and IMU data.

MosaicML released MPT-7B, Base, Instruct, Chat and StoryWriter variants under commercially usable licenses.

BigCode released StarCoder, a 15.5B-parameter code model trained on permissively licensed code from The Stack.

Baidu introduced ERNIE 3.5 as the new foundation model powering ERNIE Bot.

Meta released DINOv2, a self-supervised vision foundation model and pretrained visual features.

Amazon introduced its first text embedding foundation model through the Bedrock preview.

AWS introduced Amazon Titan, its family of text generation and embedding foundation models available through Bedrock.

Databricks released Dolly 2.0 with commercial-use weights and the human-generated databricks-dolly-15k instruction dataset.

Meta released SAM, its promptable image-segmentation model, together with model weights and the SA-1B dataset.

Databricks released Dolly, a six-billion-parameter instruction-following model that could be trained inexpensively on one machine.

Anthropic introduced Claude and Claude Instant more broadly after a closed alpha with partners.

OpenAI released GPT-4, a multimodal large model accepting image and text inputs and producing text outputs.

MosaicML released MosaicBERT, an optimized BERT architecture designed to pretrain high-quality models at much lower cost.

Cohere established Command as its principal family of enterprise generative language models.

Cohere released Command Light as a lower-latency generation option for enterprise text workloads.

Cohere introduced the Command generation family for enterprise summarization, conversation and text workflows.

Meta released the first LLaMA model family for approved research use, spanning 7B to 65B parameters.

OpenAI released text-embedding-ada-002, replacing multiple earlier embedding models with a more capable and lower-cost model.

Meta withdrew the public Galactica demo after criticism over unreliable and misleading generated scientific content.

Meta AI released a research demo of Galactica, a large language model trained on scientific knowledge.

BigScience released BLOOMZ and mT0, instruction-finetuned multilingual models derived from BLOOM and the xP3 task collection.

Cohere released its second-generation multilingual embedding model for semantic search across more than 100 languages.

Cohere established its Rerank family of cross-encoder relevance models for search pipelines.

Meta released EnCodec, a neural audio codec for high-fidelity compression.

OpenAI released Whisper, a multilingual speech-recognition model, together with model weights and inference code.

The BigScience workshop released BLOOM, a 176-billion-parameter multilingual language model trained on the Jean Zay supercomputer.

Meta released NLLB-200, a machine-translation model covering 200 languages.

Meta released the Open Pretrained Transformer family and provided OPT-175B access to researchers.

OpenAI introduced DALL-E 2 with higher-resolution image generation and editing capabilities.

OpenAI introduced instruction-following models trained with human feedback.

Cohere introduced its Embed model family for semantic representation and retrieval.

Microsoft releases DeBERTa V3 with replaced-token detection and gradient-disentangled embedding sharing.

OpenAI introduced Codex, a descendant of GPT-3 that translates natural language into code.

DeepMind publishes AlphaFold 2's methodology and releases source code, expanding access to high-accuracy protein prediction.

Google introduces LaMDA as a conversational language model designed for open-ended dialogue.

OpenAI introduced CLIP, a neural network trained on image-text pairs for transferable visual recognition.

OpenAI introduced DALL-E, a model that generates images from text descriptions.

Facebook AI released M2M-100, a multilingual machine-translation model translating directly between 100 languages.

OpenAI published GPT-3, demonstrating few-shot learning at 175 billion parameters.

OpenAI released Jukebox, a neural network that generated music including rudimentary singing as raw audio.

Microsoft introduces Turing-NLG, a 17-billion-parameter generative language model and, at the time, the largest published language model.

OpenAI completed the staged GPT-2 release with the 1.5-billion-parameter model and accompanying research.

Hugging Face released DistilBERT, a smaller and faster distilled BERT model retaining most of the teacher model’s language-understanding performance.

OpenAI released the 774-million-parameter GPT-2 model and shared further analysis of misuse and detection.

OpenAI introduced MuseNet, a deep neural network for generating multi-instrument musical compositions.

Facebook open-sourced its Deep Learning Recommendation Model reference implementation.

Baidu released its first knowledge-enhanced language representation model.

OpenAI introduced GPT-2 and began a staged release process due to misuse concerns.

Google open-sources BERT, enabling bidirectional pre-training for language understanding and downstream NLP tasks.

OpenAI presented its five-agent Dota 2 system trained through large-scale self-play.

Baidu opened PaddlePaddle to the global developer community for deep learning, vision, speech and NLP.

Google releases TensorFlow as an open-source machine-learning framework, making internal deep-learning infrastructure broadly available.

IBM publicly demonstrated its DeepQA question-answering architecture through the Jeopardy! competition.