Amazon Bedrock now supports fine-tuning for Nova models with three customization approaches

Amazon Bedrock Adds Fine-tuning for Nova Models

Amazon has announced fine-tuning capabilities for Amazon Nova models through Amazon Bedrock, enabling customers to customize models for domain-specific tasks without deep machine learning expertise.

Three Customization Approaches

Bedrock supports three fine-tuning techniques:

Supervised Fine-tuning (SFT): Trains models on labeled input-output examples, embedding domain knowledge directly into model weights.

Reinforcement Fine-tuning (RFT): Uses reward functions—either custom code or an LLM acting as judge—to guide learning toward target behaviors.

Model Distillation: Transfers knowledge from larger teacher models into smaller, faster student models for resource-constrained environments.

All three approaches use parameter-efficient fine-tuning (PEFT), reducing memory requirements and training time while maintaining model quality compared to full fine-tuning.

Supported Models

Amazon Nova 2 Lite and Nova Micro support fine-tuning. Nova 2 Lite is a multimodal model with a 1-million token context window, processing text, images, and video for document processing, video understanding, and code generation. Nova Micro, the smallest in the lineup, targets low-cost inference for pipeline processing tasks like data extraction and address fixing.

Implementation and Pricing

Amazon Bedrock automates the entire training pipeline. Users upload training data to Amazon S3 and initiate the job via AWS Management Console, CLI, or API. The service manages infrastructure provisioning, compute allocation, and training orchestration—no cluster configuration required.

Fine-tuned models run on-demand at the same inference pricing as non-customized versions, with no provisioned capacity requirement. This contrasts with traditional approaches requiring expensive Provisioned Throughput.

Performance Gains

Amazon's internal testing demonstrated measurable improvements. Amazon Customer Service customized Nova Micro for specialized support, improving accuracy by 5.4% on domain-specific issues and 7.3% on general issues while reducing latency.

Fine-tuning eliminates token consumption overhead compared to prompt engineering and Retrieval-Augmented Generation (RAG), which supply context at inference time. While context-based techniques offer immediate deployment and dynamic updates, fine-tuning embeds knowledge directly, reducing cumulative token costs and improving generalization to novel phrasings and edge cases.

When to Fine-tune

Amazon recommends fine-tuning for high-volume, well-defined tasks with quality labeled examples—such as intent classification, brand voice consistency, or replacing traditional ML classifiers. The upfront investment in data labeling and training pays off through reduced per-request inference costs for applications with sustained traffic.

Fine-tuned small LLMs like Nova Micro increasingly replace traditional classifiers for tasks requiring flexibility with natural language variation without retraining.

Training Visibility

Bedrock provides sensible hyperparameter defaults (epochCount, learningRateMultiplier) and real-time training monitoring through loss curves. Clear documentation covers data preparation, format specifications, and schema requirements.

What this means

Bedrock's fine-tuning removes infrastructure barriers for model customization, making it accessible to teams without ML ops expertise. The on-demand pricing model—eliminating provisioned capacity costs—alters economics for domain-specific deployments. This positions Nova models as viable replacements for traditional classifiers in production pipelines, particularly where cost and latency matter more than raw capability. The focus on parameter-efficient approaches preserves inference speed, critical for high-volume applications.