Generative AI platforms & models
Amazon Bedrock model customization vs Foundry fine-tuning
Bedrock model customization maps to Foundry fine-tuning.
Both offer supervised fine-tuning, reinforcement fine-tuning, and distillation for selected models.
Side-by-side comparison
| Aspect | Amazon Web Services | Microsoft Azure |
|---|---|---|
| Service | Amazon Bedrock model customization | Foundry fine-tuning |
| Lifecycle | Generally available | Generally available |
| What it is | Supervised fine-tuning, reinforcement fine-tuning (reward functions in Lambda), and distillation for supported Bedrock models. | Supervised fine-tuning, preference fine-tuning (DPO), and reinforcement fine-tuning (model-specific), plus distillation via teacher-generated data. Deploy serverless or on managed compute (LoRA adapters, preview). |
| Pricing model | Billed per training token (tokens in corpus × epochs) plus monthly model storage. Inference on a custom model may require Provisioned Throughput or on-demand custom model pricing, depending on the model. | Billed per training token or training hour (model-dependent), plus hourly hosting for deployed fine-tuned models (Developer tier available for testing). |
| Docs | Amazon Bedrock model customization docs | Foundry fine-tuning docs |
| Pricing page | Amazon Bedrock model customization pricing | Foundry fine-tuning pricing |
| Exams | Generative AI Developer – Professional MLA-C01 Retiring | AI-103 AI-300 |
Key differences
- Bedrock offers supervised fine-tuning, reinforcement fine-tuning (reward functions written in AWS Lambda, or training from invocation logs), and managed distillation (teacher to student).
- Foundry offers SFT, preference fine-tuning (DPO), and reinforcement fine-tuning (model-specific). Distillation is done through teacher-generated SFT data. Fine-tuned models deploy serverless or on managed compute with LoRA adapters (preview).
- Which base models are tunable differs a lot and changes often. Check each side's supported-model table.
Pricing summary
Pricing models only. No prices are quoted because they change by region and over time.
- AWS
- Billed per training token (tokens in corpus × epochs) plus monthly model storage. Inference on a custom model may require Provisioned Throughput or on-demand custom model pricing, depending on the model. Amazon Bedrock model customization pricing →
- Azure
- Billed per training token or training hour (model-dependent), plus hourly hosting for deployed fine-tuned models (Developer tier available for testing). Foundry fine-tuning pricing →
Certification exam mapping
| Exam | Provider | Level | Status |
|---|---|---|---|
| AWS Certified Generative AI Developer – Professional | AWS | Professional | current Exam code not confirmed from the official page at build time; shown by name only. |
| AWS Certified Machine Learning Engineer – Associate MLA-C01 | AWS | Associate | retiring Being updated to MLA-C02 (beta registration opened Sep 1, 2026). Last day for MLA-C01 in English was Sep 28, 2026; Korean, Japanese, and Simplified Chinese remain until MLA-C02 GA. |
| Microsoft Certified – Azure AI Apps and Agents Developer Associate AI-103 | Azure | Associate | current Generative AI and agentic solutions with Microsoft Foundry using Python. Successor in scope to the retired AI-102. |
| Microsoft Certified – Machine Learning Operations Engineer Associate AI-300 | Azure | Associate | current MLOps and GenAIOps infrastructure on Azure (IaC with Bicep/CLI, GitHub Actions). Closest successor in scope to the retired DP-100. |
Relevance is based on each exam's published scope, not a line-by-line objective match. Check the official study guide.
FAQ
What is the Azure equivalent of Amazon Bedrock model customization?
Bedrock model customization maps to Foundry fine-tuning. Both offer supervised fine-tuning, reinforcement fine-tuning, and distillation for selected models.
How does pricing compare between Amazon Bedrock model customization and Foundry fine-tuning?
AWS: Billed per training token (tokens in corpus × epochs) plus monthly model storage. Inference on a custom model may require Provisioned Throughput or on-demand custom model pricing, depending on the model. Azure: Billed per training token or training hour (model-dependent), plus hourly hosting for deployed fine-tuned models (Developer tier available for testing). Check the official pricing pages for current rates.
Which certification exams cover this topic?
AWS Certified Generative AI Developer – Professional; AWS Certified Machine Learning Engineer – Associate (MLA-C01) [retiring]; Microsoft Certified – Azure AI Apps and Agents Developer Associate (AI-103); Microsoft Certified – Machine Learning Operations Engineer Associate (AI-300).
Does Azure have reinforcement fine-tuning like Bedrock?
Yes, Foundry lists reinforcement fine-tuning (RFT) for supported models, alongside SFT and DPO.