Google upgrades Vertex AI with Claude 2, Llama 2, and new enterprise tools

An open mall for models

Google Cloud is not being subtle about its catch-up game. At its annual Cloud Next conference, the company rolled out a sweeping set of upgrades to Vertex AI, turning its enterprise machine learning platform into a full-blown marketplace for generative models.

On the proprietary side, Google upgraded PaLM 2 with a 32,000-token context window, roughly 80 double-spaced pages. That is still well short of Anthropic's 100,000 tokens on Claude 2, but Vertex AI product lead Nenshad Bardoliwalla pitched it as a deliberate trade-off between inference cost and context length.

Google also gave Codey a reported 25% bump in code generation quality across major languages, and updated Imagen with Style Tuning, which lets brands train image styles on as few as ten reference images.

More telling is what Google invited into its living room. Vertex AI Model Garden now officially hosts Meta's Llama 2, TII's open-source Falcon, and Anthropic's Claude 2. If customers want context sizes that PaLM 2 cannot handle, Google would rather host the competitor on its own infrastructure than lose the workload to Amazon or Microsoft.

Giving models hands and data

Raw models solve few enterprise problems on their own. To address this, Google launched Extensions and data connectors, which operate as its answer to OpenAI plugins.

Extensions allow models to take real actions on behalf of users across databases like Redis, MongoDB, and DataStax, as well as CRM and email accounts. Meanwhile, data connectors offer read-only hooks into systems like Jira, Confluence, and Salesforce.

On the application front, Google moved Vertex AI Search and Vertex AI Conversation into general availability. Formerly housed under the Generative AI App Builder label, these services now support multi-turn querying and automatic result summarization.

A new preview feature called Playbook lets builders script chatbot behaviors with plain English prompts: defining personas, explicit multi-step goals, and example transcripts rather than writing brittle workflow logic. Grounding tools also let models cite proprietary sources to rein in hallucinations, backed by safety classifiers that filter toxic outputs on API calls.

Why it matters

This update is a strategic pivot. Google spent months pushing its in-house research as the gold standard, but enterprise cloud sales require pragmatism over pride. By bringing Claude 2 and Llama 2 directly into Vertex AI, Google is taking dead aim at Amazon Bedrock, which had a rocky initial rollout.

For enterprise teams, the appeal is convenience. Having hosted third-party models alongside proprietary PaLM weights, connected to native Google IAM, VPCs, and distributed frameworks like Ray, makes Vertex a compelling one-stop shop without having to cobble together open-source serving stacks.

Still, Google left several hard questions unanswered. While it claims strict source data reviews to keep models free of copyright claims, it provides no opt-out mechanism for public web creators whose data fed these models. More pressing for legal departments, Google offered no clear stance on whether enterprise customers own copyright over the outputs their models produce. Until those legal boundaries are clarified, enterprise adoption will likely stay focused on internal workflows rather than customer-facing creative work.