Google Gemini / Vertex AI

Vertex AI Agent Engine GA with Billing Starting March 4, 2025

Google's Vertex AI Agent Engine has reached general availability and billing begins March 4, 2025. The service was previously known as LangChain on Vertex AI and has been renamed. Users should delete unused resources to avoid unwanted charges as the free preview period ends.

Date not specified

Google Colab Enterprise adds Terraform integration for notebook management

Google Colab Enterprise now supports Terraform resources for scheduling notebook runs and managing runtimes and runtime templates. This adds infrastructure-as-code capabilities for enterprise users to programmatically manage their notebook environments. No action required, this is an additive feature enhancement.

Date not specified

Google releases Gemini 2.0 Flash-Lite on Vertex AI

Google has released Gemini 2.0 Flash-Lite as a generally available model on Vertex AI. This represents a new model option for users of the platform. No immediate action is required as this is an additive capability enhancement.

Date not specified

Anthropic Claude Sonnet 3.7 now available in Preview on Vertex AI

Google has added Anthropic's Claude Sonnet 3.7 model to Vertex AI in preview mode. Users can access this new model through Model Garden with associated model card documentation. This expands the available AI model options on the Vertex AI platform.

Date not specified

Vertex AI Adds New Evaluation Metrics, Model Updates, and Enhanced Fine-tuning

Google Vertex AI introduced new evaluation metrics (perplexity, BLEU, ROUGE variants), service account selection for fine-tuning, and updated support for Llama 3.1, Gemma 2, PaliGemma 2, and LLaVA Next models. The updates include new tutorial notebooks and improved fine-tuning workflows with PEFT and Axolotl frameworks. These are additive features that enhance existing capabilities without requiring immediate action.

Date not specified

Google Vertex AI adds DeepSeek-V3 and DeepSeek-R1 models to Model Garden

Google has added two new DeepSeek models to Vertex AI Model Garden: DeepSeek-V3 (671B parameter MoE language model) and DeepSeek-R1 (671B parameter reasoning model comparable to OpenAI-o1). Both models are available in Preview mode and can be deployed using provided notebooks. This expands the available model options for developers using Vertex AI's Model Garden platform.

Date not specified

Google releases Llama 3.3 70B model in Preview on Vertex AI

Google has released the Llama 3.3 70B model in Preview status on Vertex AI's managed platform. This provides users with access to a new large language model option for their AI applications. No immediate action is required as this is a new capability addition.

Date not specified

Vertex AI Model Garden adds DeepSeek-R1 and Phi-4 models plus LLM inference optimizations

Google added two new AI models (DeepSeek-R1 and Microsoft Phi-4) to Vertex AI Model Garden and introduced advanced LLM inference optimization techniques in preview. The new optimizations include prefix caching and speculative decoding for various Llama, Gemma, and Mistral models to improve performance and reduce latency. These are additive capabilities that enhance existing services without requiring immediate action.

Date not specified

Google releases Gemini 2.0 Flash GA and new Pro/Flash-Lite models

Gemini 2.0 Flash is now generally available for text-only outputs, with multimodal capabilities remaining in private preview. Two new models have been introduced: Gemini 2.0 Pro (experimental, optimized for coding with 2M context window) and Gemini 2.0 Flash-Lite (preview, fastest and most cost-efficient option). These releases expand model options without requiring immediate user action.

Date not specified

New Monitoring Dashboard for Vertex AI Foundation Models & GA for Dedicated Endpoints

Google has released a predefined dashboard for monitoring usage, throughput, and latency of Vertex AI foundation models including Gemini and Claude. Additionally, Dedicated Public Endpoints and Private Service Connect Endpoints for Vertex AI Prediction are now generally available. These are informational updates that enhance monitoring capabilities and endpoint options without requiring immediate action.

Date not specified

Google deprecating Mistral Large and Codestral models in Vertex AI Model Garden

Google has deprecated Mistral Large (24.07) and Codestral (24.05) models offered as Model as a Service in Vertex AI Model Garden. Users currently relying on these specific model versions should plan migration to newer versions or alternative models. The deprecation timeline and replacement options are detailed in the Vertex AI deprecations documentation.

Date not specified

Google releases Imagen 3 image generation model with prompt enhancement

Google released an improved Imagen 3 model (imagen-3.0-generate-002) with enhanced image generation capabilities and automatic prompt enhancement via LLM-based rewriting. The new model is immediately available to all users and includes configurable prompt enhancement that is enabled by default to improve image quality.

Date not specified

LangChain on Vertex AI billing starts March 4, 2025

Google will begin charging for LangChain on Vertex AI usage starting March 4, 2025, based on vCPU hours ($0.0994/vCPU-Hr) and memory consumption ($0.0105/GiB-Hr). Users currently using this service should prepare for billing changes and review their usage patterns. This affects both Generative AI on Vertex AI and Vertex AI Agent Builder services.

Date not specified

Anthropic Claude 3 Sonnet Deprecated on Vertex AI Model Garden

Anthropic's Claude 3 Sonnet model offered as a Model as a Service (MaaS) in Vertex AI Model Garden has been deprecated by Google. Users currently relying on this model should plan to migrate to alternative models or providers. Check the Vertex AI deprecations documentation for migration timelines and recommended alternatives.

Date not specified

Google releases Agent evaluation service in Preview on Vertex AI

Google has released a new Agent evaluation feature using the Gen AI evaluation service in Preview status on Vertex AI. This allows developers to evaluate AI agents during development. No immediate action required as this is a new optional feature in preview.

Date not specified

Vertex AI Workbench M127 Release Fixes SSH Key Home Directory Ownership Issue

The M127 release addresses a bug related to home directory ownership when using authorized SSH keys across all Workbench notebook types. This is a maintenance fix that improves SSH authentication reliability. No user action is required as this is an automatic improvement to existing functionality.

Date not specified