Amazon Bedrock (model customization)
Amazon Web Services, Inc. · aws.amazon.com · Last checked:
- Managed fine-tuning API
- Cloud platform fine-tuning
- Reinforcement fine-tuning
Company status
Active. Amazon Bedrock is a business line of Amazon Web Services; parent Amazon.com, Inc. is publicly listed.
Sources 1, 2, 3, 4 · checked 2026-11-08
Capabilities
- Supervised fine-tuning on labeled data for supported Amazon Nova, Meta Llama 3.x, and Anthropic Claude 3 Haiku models
- Reinforcement fine-tuning with reward functions defined in AWS Lambda; can use prompt datasets or Bedrock invocation logs
- Distillation from a larger teacher model to a smaller student model
- Custom models are used inside Bedrock; training billed by tokens processed plus monthly model storage
Sources 1, 2, 3, 4 · checked 2026-11-08
Who it tends to fit
- Teams already on AWS that want customization inside their existing account and controls
- Buyers who need HIPAA-eligible or FedRAMP High (GovCloud US-West) services
Sources 1, 2, 3, 4 · checked 2026-11-08
Pricing
Reinforcement fine-tuning: $80.00 per training hour for gpt-oss-20b and Qwen3 32B (US East N. Virginia and US West Oregon), plus $1.95 storage per trained model per month. Supervised fine-tuning is billed per token processed (training tokens x epochs) plus monthly storage; rates vary by model.
Sources 1, 2, 3, 4 · checked 2026-11-08
Published certifications
- In scope for AWS ISO and SOC compliance programs (as stated by AWS)
- HIPAA eligible
- GDPR (customers can use Bedrock in compliance with GDPR, per AWS)
- FedRAMP High (AWS GovCloud US-West)
Listed as stated by the vendor. Scope can differ by product or subsidiary; request current reports.
Sources 1, 2, 3, 4 · checked 2026-11-08
Limitations to check
- Fine-tuning model list is limited and region-specific (for example, Nova models in us-east-1, Llama models in us-west-2)
- No weights-export path found in the Bedrock customization docs; custom models run inside Bedrock
- Supervised fine-tuning rates for current models were not readable on the public pricing page when checked; confirm in the AWS pricing calculator
Sources 1, 2, 3, 4 · checked 2026-11-08
Sources
- Customization methods: supervised fine-tuning, reinforcement fine-tuning (Lambda reward functions), distillation; training charged by tokens x epochs plus monthly storagedocs.aws.amazon.com/bedrock/latest/userguide/custom-models.htmlChecked 2026-11-08
- Fine-tunable models and regions table (Nova, Llama 3.x, Claude 3 Haiku)docs.aws.amazon.com/bedrock/latest/userguide/custom-model-fine-tuning.htmlChecked 2026-11-08
- RFT pricing $80.00 per training hour for gpt-oss-20b and Qwen3 32B; storage $1.95 per trained model per monthaws.amazon.com/bedrock/pricing/Checked 2026-11-08
- Bedrock in scope for ISO, SOC, CSA STAR Level 2; HIPAA eligible; GDPR; FedRAMP High in GovCloud (US-West)aws.amazon.com/bedrock/security-compliance/Checked 2026-11-08
Report a correction: [email protected]. Facts are re-checked at least every 90 days.