Amazon SageMaker AI now supports serverless model customization for Gemma 4 models
Fine-tune Gemma 4 on SageMaker AI with no cluster management — pay only for training time using SFT, DPO, or RFT.
View original announcement →Visual Summary
graph TD
A{{SageMaker AI Serverless Model Customization}}:::announced
B((AI Developer)):::external
C(SageMaker Studio):::compute
D(SageMaker Python SDK V3):::compute
E([Gemma 4 Models]):::feature
F([SFT / DPO / RFT]):::feature
G(GPU Infrastructure):::compute
H(Proprietary Data):::storage
I([Built-in Evaluation]):::feature
J(SageMaker Endpoints):::compute
B ==>|"launches job"| A
B -->|"via UI"| C
B -->|"via SDK"| D
C ==>|"configures"| A
D ==>|"configures"| A
H -->|"training data"| A
A -->|"selects"| E
A -->|"applies"| F
A -->|"provisions"| G
A -->|"evaluates"| I
A -.->|"deploys to"| J
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
Amazon SageMaker AI now supports serverless model customization for Google DeepMind's Gemma 4 E4B and 31B models, enabling fine-tuning via supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement fine-tuning (RFT). This launch also broadens the overall catalog of models eligible for serverless customization on SageMaker AI, adding Gemma 4 alongside existing support for Nova, Nemotron 3, Qwen, Llama, gpt-oss, and DeepSeek families. Users can launch customization jobs directly from SageMaker Studio or programmatically via the SageMaker Python SDK V3.
How It Works
- Model selection: Users choose a supported foundation model (e.g., Gemma 4 E4B or 31B) from the Models page in SageMaker Studio or via the SageMaker Python SDK V3, which is required for this feature.
- Customization technique selection: Three training paradigms are supported — SFT (supervised fine-tuning on labeled input/output pairs), DPO (direct preference optimization for aligning outputs to human preferences), and RFT (reinforcement fine-tuning using verifiable reward signals).
- Serverless infrastructure provisioning: SageMaker AI automatically provisions GPU instances (P5, P4de, P4d, or G5) sized appropriately for the chosen model and technique, with no manual cluster configuration required.
- Pre-optimized training recipes: The platform applies best-practice training configurations for each model-technique combination, reducing the need for hyperparameter expertise.
- Real-time monitoring: Live metrics and logs are surfaced through the SageMaker Studio UI during training, enabling users to track job progress without managing underlying infrastructure.
- Automatic resource cleanup: After training completes, all provisioned compute resources are automatically terminated, preventing idle cost accumulation.
- Evaluation and deployment integration: The SDK V3 includes built-in evaluation tools (LLM-as-Judge, InspectAI, custom scorers) and supports direct deployment to SageMaker endpoints or Amazon Bedrock via BedrockModelBuilder.
Why It's Important
- Access to a leading open model family: Gemma 4 models from Google DeepMind are competitive open-weight models; serverless fine-tuning support makes them practically accessible for enterprise customization without infrastructure expertise.
- Reduces time-to-production: The announcement explicitly positions this as compressing what was previously a months-long process into days, directly accelerating AI product development cycles.
- Lowers the barrier to advanced alignment techniques: DPO and RFT are sophisticated alignment methods that previously required significant ML engineering effort; managed support democratizes access to these techniques.
- Cost efficiency: The pay-per-use serverless model eliminates the need to reserve or manage GPU clusters, which is especially valuable for teams running infrequent or experimental fine-tuning jobs.
- Expanding model diversity: The simultaneous expansion of the serverless catalog (Nova, Nemotron 3, Qwen, Llama, gpt-oss, DeepSeek, and now Gemma 4) gives organizations flexibility to fine-tune the model best suited to their use case without switching platforms.
- Proprietary data utilization: Organizations can inject domain-specific knowledge, tone, and task-specific behavior into powerful base models using their own labeled datasets, creating differentiated AI assets.
How It's Different
- Serverless vs. self-managed fine-tuning: Unlike traditional SageMaker Training Jobs where users must configure instance types, distributed training frameworks, and container images, serverless customization abstracts all of this away entirely.
- Guided UI vs. code-only workflows: The SageMaker Studio interface provides a natural-language-aware guided experience for launching jobs, contrasting with raw notebook or CLI-based fine-tuning pipelines common in open-source tooling.
- Integrated evaluation vs. separate tooling: Built-in evaluators (LLM-as-Judge, benchmark scoring, custom scorers) are co-located with the training workflow, whereas most fine-tuning platforms require separate evaluation infrastructure.
- Multi-technique support in one platform: SFT, DPO, and RFT are all supported natively, whereas many managed fine-tuning services (including Amazon Bedrock's native fine-tuning) support only SFT or a limited subset of techniques.
- Automatic resource cleanup vs. persistent clusters: Resources are torn down post-training automatically, unlike self-managed GPU clusters that accrue cost when idle.
- SDK V3 exclusivity: This capability is only available in SageMaker Python SDK V3, which was purpose-built for foundation model workflows, distinguishing it from legacy SageMaker training APIs.
When to Prefer It
- Domain adaptation without ML infrastructure teams: Organizations that have domain-specific data but lack dedicated MLOps or GPU cluster management expertise should prefer this over self-managed fine-tuning.
- Preference alignment use cases: When outputs need to reflect organizational tone, safety guidelines, or ranked human preferences, DPO is the appropriate technique and is fully managed here.
- Reasoning and verifiable task improvement: For use cases like math, code generation, or structured output tasks where correctness can be verified, RFT provides reward-signal-based training that is otherwise complex to implement.
- Infrequent or experimental fine-tuning: Teams that run fine-tuning jobs occasionally (e.g., monthly model refreshes or A/B experiments) benefit most from pay-per-use serverless pricing versus reserved capacity.
- Gemma 4 as a preferred base model: Organizations already evaluating or deploying Gemma 4 for its architecture or licensing characteristics can now fine-tune it in the same platform used for inference.
- Rapid prototyping to production pipelines: When the goal is to move from a proof-of-concept fine-tuned model to a deployed endpoint quickly, the integrated evaluation and deployment capabilities reduce handoff friction.
- Multi-model strategy: Enterprises comparing fine-tuned variants across model families (e.g., Gemma 4 vs. Llama vs. Qwen) can run and evaluate all jobs within a single platform for consistent comparison.
Availability
- Status: Generally available (GA) as of June 30, 2026.
- Supported regions: US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and EU (Ireland).
- Pricing model: Serverless, pay-per-use — charges apply only for compute consumed during active training jobs; no idle cluster costs.
- Supported models for Gemma 4: Gemma 4 E4B and Gemma 4 31B parameter variants.
- Supported customization techniques: Supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement fine-tuning (RFT).
- SDK requirement: Serverless model customization requires SageMaker Python SDK V3; it is not available in earlier SDK versions.
- Access methods: SageMaker Studio (Models page UI) or SageMaker Python SDK V3 for programmatic access.
- Compute: Infrastructure automatically selects from P5, P4de, P4d, and G5 GPU instance families based on model size and technique requirements.
