Releases

DeepSeek published DeepSeek-V2 base and chat weights with 128K context.

DeepSeek introduced its efficient second-generation mixture-of-experts architecture.

Amazon released its second-generation multilingual text embedding model through Bedrock.

Microsoft releases Phi-3 Mini, a 3.8-billion-parameter model designed for capable language and reasoning on constrained devices.

Synthesia released EXPRESS-1, its fourth-generation avatar model for context-aware voice, facial expression, gaze, and body language.

Meta and World Resources Institute released global high-resolution canopy-height maps built with DINOv2.

Meta released the Llama 3-based update to its open safeguard model.

Meta released the 8B Llama 3 pretrained and instruction-tuned models.

Meta released the 70B Llama 3 pretrained and instruction-tuned models.

Hugging Face released IDEFICS2, an 8B visual-language model with stronger OCR, document, and visual reasoning capabilities.

Mistral AI released Mixtral 8x22B, a larger open sparse mixture-of-experts model.

xAI previewed its first multimodal model, adding understanding of documents, diagrams, charts, screenshots and photographs.

Google releases CodeGemma variants for code completion, code generation and instruction following.

Databricks optimized DBRX inference and made the model available through Mosaic AI Model Serving.

Cohere introduced Command R+, its higher-capacity enterprise model for advanced RAG, multilingual generation and tools.

Replit introduced a native 7B model specialized for fixing language-server diagnostic errors.

xAI announced Grok-1.5 with improved reasoning and a 128,000-token context window.

Databricks released DBRX, a 132-billion-total-parameter mixture-of-experts model with 36 billion active parameters under an open license.

Cohere Labs made Command R weights available for research and evaluation.

Cohere released Command R, a 128K-context model optimized for production RAG, citations and tool use.

Suno released v3 with improved full-song generation, vocals, and instrumentation.

NVIDIA announced a general-purpose foundation model initiative for humanoid robots.

xAI released the 314B-parameter Grok-1 base weights and Mixture-of-Experts architecture under Apache 2.0.

Anthropic released Claude 3 Haiku as the fastest and least expensive model in the Claude 3 family.

Amazon researchers released pretrained Chronos-T5 time-series foundation models and code.

DeepSeek released 1.3B and 7B base and chat checkpoints.

SambaNova announced Samba-1, a 1.3-trillion-parameter Composition of Experts system comprising 92 specialist models behind one endpoint.

Claude 3 Opus and Claude 3 Sonnet became available on Claude.ai and the generally available Claude API.

Anthropic announced Claude 3 Opus, Sonnet, and Haiku, adding multimodal image understanding and a three-tier capability lineup.

Harvey and OpenAI developed a custom domain-specific model for legal case-law tasks.

The first proprietary Tab model established Cursor's in-house coding-model family.

Cursor shipped its first custom sparse language model for edit prediction, powering Cursor Tab suggestions.

BigCode released StarCoder2 models at 3B, 7B, and 15B scales in collaboration with ServiceNow and NVIDIA.

Mistral AI introduced Mistral Small as its efficient commercial model tier.

Mistral AI launched its first flagship commercial model, Mistral Large.

Google releases Gemma 2B and 7B open models with weights and responsible AI tooling.

Meta released V-JEPA, a video model trained by predicting missing representations rather than pixels.

Google DeepMind introduces Gemini 1.5 Pro with a mixture-of-experts architecture and a one-million-token context window.

OpenAI introduced Sora, a text-to-video model, as a research preview for red teamers and selected creators.

Cohere For AI released Aya 101, an open multilingual instruction-following model covering 101 languages.

Gemini 1.0 Ultra becomes available to consumers through the new Gemini Advanced subscription.

Base, instruction and reinforcement-learning DeepSeekMath 7B checkpoints were published.

Midjourney and Spellbrush released Niji V6.

Meta released a 70-billion-parameter Code Llama model for code generation and instruction following.

DeepSeek released an updated 7B coding model that later served as the initialization for DeepSeekMath.

OpenAI released text-embedding-3-small and text-embedding-3-large with improved performance and new dimension controls.

Oracle made Meta Llama 2 available as a managed model in OCI Generative AI.

Oracle made Cohere generation and embedding models available through its managed service.

Midjourney released V6 with better prompt accuracy, text rendering, coherence, and image prompting.

Google makes Imagen 2 available to Cloud customers through Vertex AI for enterprise image generation.