Vertex AI Tuning

source

closed

training-code

closed

runs-on

Google-Cloud-Vertex-AI-only

license

Proprietary(managed GCP service)

Proprietary managed tuning service inside Google Cloud's Vertex AI generative-AI docs (now also mirrored under 'Gemini Enterprise Agent Platform' branding); no source, runs only on Google Cloud against closed Gemini models.

• https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/tune-models recorded 2026-08-09

  The tuning introduction page (redirected today to the canonical docs.cloud.google.com/gemini-enterprise-agent-platform/models/tuning) describes Gemini tuning entirely as a hosted capability: supervised fine-tuning, preference tuning, tuning checkpoints and continuous tuning are all run and served by Google, opening with "Agent Platform supports supervised fine-tuning to customize foundational models." No client library, SDK source, or self-hostable runtime is offered anywhere on the page.

• https://cloud.google.com/terms/ recorded 2026-08-09

  Section 3.3 "Restrictions" of the Google Cloud Platform Terms of Service states Customer will not, and will not allow End Users to, "copy, modify, or create a derivative work of the Services", "reverse engineer, decompile, translate, disassemble, or otherwise attempt to extract any or all of the source code of, the Services", or "sell, resell, sublicense, transfer, or distribute any or all of the Services" — the standard proprietary-SaaS terms this closed, managed feature is offered under.

No usage figure is disclosed for Gemini tuning. The pricing page publishes only per-token tuning rates, and the Gemini app's monthly active users belong to a different product and cannot be attributed to this service, so no level is assigned rather than one borrowed from that surface.

• https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/tune-models recorded 2026-08-09

  The tuning documentation is entirely a how-to and reference guide; it contains no count of customers, developers, tuning jobs, or tokens tuned anywhere on the page.

• https://cloud.google.com/vertex-ai/generative-ai/pricing recorded 2026-08-09

  The pricing page (redirected today to cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing) lists only per-token rates, e.g. a 'Tuning for 1M training tokens' row at $3.00 for Gemini 2.0 Flash and $1.00 for Gemini 2.0 Flash-Lite. A case-insensitive search of the extracted page text for 'customers', 'users', 'downloads', 'MAU' and 'monthly active' turns up pricing-terms usage and case-study links only, no adoption or usage-count figure for the tuning feature.

Broad managed multimodal tuning over frontier closed Gemini models: supervised fine-tuning, preference tuning in the DPO mold and a preview reinforcement-learning method, applied across text, documents, images, audio and video, plus function-calling tuning and checkpoint management. It is a black box, with no scale or precision controls and no benchmark result behind it, so it is capped below the open-source libraries that define scale in this category; the score reflects breadth of method and modality.

• https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/tune-models recorded 2026-08-09

  Confirms three tuning methods for Gemini models: "Supervised fine-tuning"; "Preference tuning", described as "Gemini Enterprise Agent Platform preference tuning builds on supervised fine-tuning by letting you tune your Gemini models with human feedback data"; and a nav-listed "Reinforcement learning fine-tuning" section carrying a "Preview" badge (data-title="Preview"). Also lists "Tuning checkpoints" and "Continuous tuning". The 'Supported modalities' nav enumerates Text, Document, Image, Audio and Video tuning plus "Tune function calling".