Skip to main content
Watch: I Built an AI App Without Leaving Databricks

Fine-tuning - verified

Google Vertex AI fine-tuning: status, methods and prices

Vertex AI only. Fine-tuning is not offered in the Gemini API or AI Studio; 4 offerings remain on Vertex AI. 4 offerings recorded: SFT, DPO and RFT. Training is priced per 1M training tokens and per 1,000 training tokens, in Google's own units. Weight export: Not stated. Google's own pages disagree on 1 point; both versions are quoted below.

Recorded as: Google (Gemini API and Vertex AI / Gemini Enterprise Agent Platform).

Read from docs.cloud.google.com, cloud.google.com, ai.google.dev. Compare every fine-tuning API

Status
Vertex AI only
Offerings
4
Methods
SFT, DPO, RFT
Export weights
Not stated

Status

Is Google Vertex AI fine-tuning still available?

Vertex AI only

Fine-tuning is not offered in the Gemini API or AI Studio; 4 offerings remain on Vertex AI.

“With the deprecation of Gemini 1.5 Flash-001 in May 2025, we no longer have a model available which supports fine-tuning in the Gemini API or AI Studio”
Gemini API docs - ai.google.dev

Dated changes

  1. Passed

    Gemini API fine-tuning ended with Gemini 1.5 Flash-001 deprecation (May 2025)

    “With the deprecation of Gemini 1.5 Flash-001 in May 2025, we no longer have a model available which supports fine-tuning in the Gemini API or AI Studio”
    ai.google.dev

The docs disagree

Google Vertex AI's own pages say different things here. Both are quoted; neither is picked.

The docs disagree

Whether Gemini 3.5 Flash Lite can be tuned

The Vertex AI pricing table lists a supervised fine-tuning price for Gemini 3.5 Flash Lite, but the supervised tuning page's list of supported Gemini models does not include it.

“Gemini 3.5 Flash Lite · Supervised fine-tuning · $0.003 / 1,000 count”
Vertex AI pricing table - cloud.google.com
“The following Gemini models support supervised fine-tuning:”
Supervised tuning page (lists Gemini 3.5 Flash, 3.1 Flash-Lite, 2.5 Pro, 2.5 Flash-Lite and 2.5 Flash only) - docs.cloud.google.com

We quote both and do not pick one. Ask the provider which applies to your account.

All 4 offerings

Prices are in the provider's own unit, as written. Not stated means the page we checked does not say.
SFTAvailableInferredDocs disagree

Vertex AI supervised fine-tuning (Gemini)

Supervised fine-tuning

Status

Available (per-model GA vs preview not stated in page text)

Limitation-table anchors for Gemini 3.5 Flash and Gemini 3.1 Flash-Lite are named '...-public-preview', suggesting those two are Preview; not stated in visible text.

Inferred: the page does not state this outright.

“The following Gemini models support supervised fine-tuning:”
docs.cloud.google.com
Tunable base models
  • Gemini 3.5 Flash
  • Gemini 3.1 Flash-Lite
  • Gemini 2.5 Pro
  • Gemini 2.5 Flash-Lite
  • Gemini 2.5 Flash

Display names, not API model IDs; the page lists names only.

“Gemini 3.5 Flash”
docs.cloud.google.com
“Gemini 3.1 Flash-Lite”
docs.cloud.google.com
“Gemini 2.5 Pro”
docs.cloud.google.com
“Gemini 2.5 Flash-Lite”
docs.cloud.google.com
“Gemini 2.5 Flash”
docs.cloud.google.com
Training price
Vertex AI supervised fine-tuning (Gemini) training prices in Google (Gemini API and Vertex AI / Gemini Enterprise Agent Platform)'s own units
ModelPriceUnitQuote
Gemini 3.5 Flash$0.01per 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 3.5 Flash · Supervised fine-tuning, Reinforcement Learning fine-tuning · $0.01 / 1,000 count”cloud.google.com
Gemini 3.5 Flash Lite$0.003Docs disagreeper 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 3.5 Flash Lite · Supervised fine-tuning · $0.003 / 1,000 count”cloud.google.com
Gemini 3.1 Flash Lite$0.003per 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 3.1 Flash Lite · Supervised fine-tuning · $0.003 / 1,000 count”cloud.google.com
Gemini 2.5 Pro$0.025per 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 2.5 Pro · Supervised fine-tuning · $0.025 / 1,000 count”cloud.google.com
Gemini 2.5 Flash$0.005per 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 2.5 Flash · Supervised fine-tuning, Preference tuning · $0.005 / 1,000 count”cloud.google.com
Gemini 2.5 Flash Lite$0.0015per 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 2.5 Flash Lite · Supervised fine-tuning, Preference tuning · $0.0015 / 1,000 count”cloud.google.com
Tuned-model inference

  • Gemini 3+ tuned endpoints: 1.5x base-model prediction price; older Gemini: same as base

    “For model inference starting from Gemini 3, tuned model endpoint prediction price will be 1.5 times of the base model. Old Gemini models prediction price stays the same as the base model.”
    cloud.google.com
Hosting requirement

tuned models served from 'us' and 'eu' multi-region endpoints only

“Supported endpoint for tuned model serving · us and eu multi-region endpoints, only”
docs.cloud.google.com
Export weights

Not stated

Data limits

  • max 1GB JSONL training file

    “Maximum training dataset file size · 1GB for JSONL”
    docs.cloud.google.com

  • max 10M text-only or 300K multimodal examples

    “Maximum training dataset size · 10M text-only examples or 300K multimodal examples”
    docs.cloud.google.com

  • max 131,072 input+output tokens per training example

    “Maximum input and output tokens per training example · 131,072”
    docs.cloud.google.com
Notes

Pricing table also lists 'Gemini 3.5 Flash Lite' for SFT, which the supported-models list on the tuning page does not include. Tuning regions: us-central1 and europe-west4.

DPONot stated

Vertex AI preference tuning (Gemini)

Direct preference optimization (preference tuning)

Status
Not stated
Tunable base models
  • Gemini 2.5 Flash
  • Gemini 2.5 Flash Lite
“Gemini 2.5 Flash · Supervised fine-tuning, Preference tuning”
cloud.google.com
“Gemini 2.5 Flash Lite · Supervised fine-tuning, Preference tuning”
cloud.google.com
Training price
Vertex AI preference tuning (Gemini) training prices in Google (Gemini API and Vertex AI / Gemini Enterprise Agent Platform)'s own units
ModelPriceUnitQuote
Gemini 2.5 Flash$0.005per 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 2.5 Flash · Supervised fine-tuning, Preference tuning · $0.005 / 1,000 count”cloud.google.com
Gemini 2.5 Flash Lite$0.0015per 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 2.5 Flash Lite · Supervised fine-tuning, Preference tuning · $0.0015 / 1,000 count”cloud.google.com
Tuned-model inference

  • Gemini 3+ tuned endpoints: 1.5x base-model prediction price; older Gemini: same as base

    “For model inference starting from Gemini 3, tuned model endpoint prediction price will be 1.5 times of the base model. Old Gemini models prediction price stays the same as the base model.”
    cloud.google.com
Hosting requirement
Not stated
Export weights

Not stated

Data limits
Not stated
Notes

Evidence is the pricing table only; the preference-tuning docs page was not fetched, so status and data limits are not recorded.

RFTNot stated

Vertex AI reinforcement learning fine-tuning (Gemini)

Reinforcement fine-tuning

Status
Not stated
Tunable base models
  • Gemini 3.5 Flash
“Gemini 3.5 Flash · Supervised fine-tuning, Reinforcement Learning fine-tuning”
cloud.google.com
Training price
Vertex AI reinforcement learning fine-tuning (Gemini) training prices in Google (Gemini API and Vertex AI / Gemini Enterprise Agent Platform)'s own units
ModelPriceUnitQuote
Gemini 3.5 Flash$0.01per 1,000 training tokens (page unit: '1,000 count'; training tokens = dataset tokens x epochs)“Gemini 3.5 Flash · Supervised fine-tuning, Reinforcement Learning fine-tuning · $0.01 / 1,000 count”cloud.google.com
Tuned-model inference

  • Gemini 3+ tuned endpoints: 1.5x base-model prediction price; older Gemini: same as base

    “For model inference starting from Gemini 3, tuned model endpoint prediction price will be 1.5 times of the base model. Old Gemini models prediction price stays the same as the base model.”
    cloud.google.com
Hosting requirement
Not stated
Export weights

Not stated

Data limits
Not stated
Notes

Evidence is the pricing table only; the RL tuning docs page was not fetched, so status and data limits are not recorded.

SFTNot stated

Vertex AI supervised fine-tuning (open models: Gemma, Llama, Qwen)

Supervised fine-tuning

Status
Not stated
Tunable base models
  • Gemma 3 1B IT
  • Gemma 3 4B IT
  • Gemma 3 12B IT
  • Gemma 3 27B IT
  • Medgemma 1.5 4B IT
  • Llama 3.1 8B
  • Llama 3.2 1B
  • Llama 3.2 3B
  • Llama 3.3 70B
  • Llama 4 Scout 17B 16E
  • Qwen 3 4B
  • Qwen 3 8B
  • Qwen 3 14B
  • Qwen 3 32B

Full list taken from the same pricing table; three representative rows quoted.

“Gemma 3 27B IT · Supervised fine-tuning · $6.83 / 1,000,000 count”
cloud.google.com
“Llama 3.3 70B · Supervised fine-tuning · $6.72 / 1,000,000 count”
cloud.google.com
“Qwen 3 32B · Supervised fine-tuning · $6.57 / 1,000,000 count”
cloud.google.com
Training price
Vertex AI supervised fine-tuning (open models: Gemma, Llama, Qwen) training prices in Google (Gemini API and Vertex AI / Gemini Enterprise Agent Platform)'s own units
ModelPriceUnitQuote
Gemma 3 1B IT$0.47per 1,000,000 training tokens (page unit: '1,000,000 count')“Gemma 3 1B IT · Supervised fine-tuning · $0.47 / 1,000,000 count”cloud.google.com
Gemma 3 27B IT$6.83per 1,000,000 training tokens (page unit: '1,000,000 count')“Gemma 3 27B IT · Supervised fine-tuning · $6.83 / 1,000,000 count”cloud.google.com
Llama 3.1 8B$0.67per 1,000,000 training tokens (page unit: '1,000,000 count')“Llama 3.1 8B · Supervised fine-tuning · $0.67 / 1,000,000 count”cloud.google.com
Llama 3.3 70B$6.72per 1,000,000 training tokens (page unit: '1,000,000 count')“Llama 3.3 70B · Supervised fine-tuning · $6.72 / 1,000,000 count”cloud.google.com
Llama 4 Scout 17B 16E$5.77per 1,000,000 training tokens (page unit: '1,000,000 count')“Llama 4 Scout 17B 16E · Supervised fine-tuning · $5.77 / 1,000,000 count”cloud.google.com
Qwen 3 8B$4.18per 1,000,000 training tokens (page unit: '1,000,000 count')“Qwen 3 8B · Supervised fine-tuning · $4.18 / 1,000,000 count”cloud.google.com
Qwen 3 32B$6.57per 1,000,000 training tokens (page unit: '1,000,000 count')“Qwen 3 32B · Supervised fine-tuning · $6.57 / 1,000,000 count”cloud.google.com
Tuned-model inference
Not stated
Hosting requirement
Not stated
Export weights

Not stated

Data limits
Not stated
Notes

Open-model tuning docs pages (hosting, weight export) were not fetched; only pricing is recorded.

Research notes

Gemini API / AI Studio: no tunable model since Gemini 1.5 Flash-001 was deprecated (May 2025); tuning lives in Vertex AI, whose docs now brand it 'Gemini Enterprise Agent Platform' (vertex-ai URLs redirect there). Vertex supervised tuning is adapter-based (adapter sizes 1-16). Supervised fine-tuning is excluded from SLAs.

Plain language

The methods, explained

SFTSupervised fine-tuning
You show the model example prompts with the answers you want, and it learns to answer like that. The default method almost everywhere.
DPODirect preference optimization (preference tuning)
You give pairs of answers to the same prompt, one preferred and one not, and the model learns to lean toward the preferred kind. Good for tone and style.
RFTReinforcement fine-tuning
The model attempts tasks and a grader (your code, or another model) scores each attempt; training pushes it toward higher-scoring answers. Often billed by training hour, plus the grader's tokens.
Vision SFTVision fine-tuning
Supervised fine-tuning where the examples include images, so the model gets better at a visual task.
DistillationDistillation
A large teacher model writes answers that are used to fine-tune a smaller, cheaper student model. You pay for the teacher's calls as well as the training.
LoRALoRA (adapter) fine-tuning
Trains a small set of extra weights, an adapter, on top of the frozen base model. Cheaper and faster than full fine-tuning; the adapter is served alongside the base model.
FullFull-parameter fine-tuning
Updates every weight in the model. Costs more than LoRA and produces a complete copy of the model.
CPTContinued pre-training
Keeps training the model on a large body of raw, unlabeled text from your domain, so it absorbs vocabulary and facts before any instruction tuning.

FAQ

Does Google offer fine-tuning?

Offered on Vertex AI; no tunable model in the Gemini API. Fine-tuning is not offered in the Gemini API or AI Studio; 4 offerings remain on Vertex AI. The source says: "With the deprecation of Gemini 1.5 Flash-001 in May 2025, we no longer have a model available which supports fine-tuning in the Gemini API or AI Studio"

How much does Google fine-tuning cost?

Per 1M training tokens: Gemma 3 1B IT $0.47, Gemma 3 27B IT $6.83, Llama 3.1 8B $0.67, Llama 3.3 70B $6.72 and 3 more. Per 1,000 training tokens: Gemini 3.5 Flash $0.01, Gemini 3.5 Flash Lite $0.003, Gemini 3.1 Flash Lite $0.003, Gemini 2.5 Pro $0.025 and 5 more. Prices are in Google's own units and are not converted.

Can you download a model you fine-tune on Google?

Not stated. The pages we checked do not say whether tuned weights can be exported.

Sources

Last verified . This is a directory of what each provider's own pages say, not a quality ranking. Prices are copied in the provider's own unit and never converted, so rows in different units are not comparable.

Newsletter

Know when a provider changes its fine-tuning terms

One email per week. New model releases, deprecations and pricing changes, and what they mean for what you build.