Vertex Artificial Intelligence release notes

A chronological log of production updates for Vertex Artificial Intelligence on Google Cloud, covering new models, platform features, deprecations, security notices, and tooling changes. The page is maintained as the authoritative source for feature launches and lifecycle changes through November 13, 2025.

This page documents production updates to Vertex Artificial Intelligence, including feature launches, changes, deprecations, security bulletins, and bug fixes. The release notes span multiple product areas such as Model Garden managed models, Vertex Artificial Intelligence Agent Engine, image and video generation models, grounding and retrieval features, and SDK/tooling releases. The document is updated regularly and was last updated on 2025-11-13 UTC.

Highlights include major model additions and lifecycle changes in Model Garden. Google and third-party models mentioned include Gemini 2.5 Flash, Gemini 2.5 Pro, Gemini 2.5 Flash-Lite, Imagen 4, MiniMax M2, DeepSeek variants, Qwen3-VL, Earth AI, MedGemma, Gemma 3, Llama 3 variants, and Mistral Codestral 2. Imagen 4 reached general availability and several Imagen preview models and older Imagen versions were scheduled for deprecation. Video model Veo progressed across versions (Veo 2, Veo 3, Veo 3.1) with features such as object insertion/removal, upscaling and short-duration generation. New and updated embedding models introduced dynamic embedding sizes and stable embedding releases. Model Garden also gained advanced inference optimizations in Preview such as prefix caching and speculative decoding, and new serving options like vLLM TPU and Hex-LLM for high-efficiency serving.

Platform and tooling changes are prominent: Vertex Artificial Intelligence Agent Engine moved to general availability with a sequence of Preview and GA features for observability, memory bank, code execution, agent-to-agent protocols, playground testing, evaluation integration, IAM-based agent identities, express mode, and billing updates. Prompt management and prompt sharing were added to Vertex Artificial Intelligence Studio, the Vertex prompt optimizer reached GA, and the Google Gen Artificial Intelligence SDK arrived in C# Preview. Grounding features evolved (including grounding with Google Maps and Google Search), RAG Engine reached GA, context caching for Gemini went GA, and the Live API and Gemini Live Native Audio arrived in Preview with native audio and improved barge-in and transcription. The notes also record deprecations and a resolved security incident affecting streaming for certain third-party models; Anthropic Claude 3.7 Sonnet and other models were listed for deprecation with scheduled removal dates. Readers are directed to the feed URL and console release notes for programmatic access and subscription to updates.

55

Impact Score

Teaching large language models how to absorb new knowledge

Researchers at MIT have developed a self-adapting framework that lets large language models permanently internalize new information by generating and learning from their own self-edits. The method could help Artificial Intelligence agents update between conversations and adapt to changing tasks.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.