Amazon Bedrock Custom Models / Fine-Tuning
Amazon Web ServicesAmazon Bedrock Custom Models is AWS's managed fine-tuning service for Bedrock foundation models, covering supervised fine-tuning, reinforcement fine-tuning, and distillation across Claude 3 Haiku, Meta Llama, and Amazon Titan and Nova. Reinforcement fine-tuning grades responses through AWS Lambda reward functions rather than labeled examples, and a customized model runs only under purchased Provisioned Throughput.
AWS states Bedrock is the only fully managed service that can fine-tune Claude models; that is the vendor's own claim, not an independent check. Verified 2026-08-09 via the AWS Bedrock model-customization documentation and AWS's Claude 3 Haiku fine-tuning GA announcement.
Openness
1 high confidence- source
- closed
- training-code
- closed
- runs-on
- AWS-Bedrock-only
- license
- Proprietary(managed AWS service, token+storage billed)
Proprietary managed customization service inside AWS Bedrock, no source, runs only on AWS against Bedrock-hosted foundation models.
- https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html recorded 2026-08-09
Lists supervised fine-tuning, reinforcement fine-tuning, and distillation as the three customization methods, delivered entirely inside Bedrock via console/API with no repository, download, or self-hostable engine referenced anywhere on the page. A billing note states training is charged by tokens processed "and model storage charged per month per model," a metered AWS commercial arrangement rather than any code license -- consistent with the recorded source:closed and license:Proprietary components.
Adoption
not assessedNo usage figure is disclosed for Bedrock's customization feature -- no jobs run, no count of customers fine-tuning. Bedrock as a whole is widely used, but nothing published is specific to the custom-models and fine-tuning capability, so no level is assigned.
- https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html recorded 2026-08-09
Same page, fetched today, lists the three customization methods and the token/storage billing note but discloses no usage figure for the customization or fine-tuning capability specifically -- no count of jobs run, models customized, or customers using it, and nothing on the page substitutes a number. The null level stands as a deliberate abstention rather than an invented band.
Capability
3 medium confidenceA solid managed method set: supervised fine-tuning, reinforcement fine-tuning with custom Lambda reward functions and real-time metrics, teacher-to-student distillation, and continued pretraining on selected models. It remains a black-box service with no exposed parallelism, precision or scale controls and no training benchmark behind it. Capable fine-tuning across a model catalog, then, but below the reinforcement-fine-tuning-on-frontier-reasoning depth of the OpenAI and Vertex offerings, and far below the open-source libraries that define scale here.
- https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html recorded 2026-08-13
SFT, reinforcement fine-tuning with Lambda reward funcs, distillation
Verified 2026-08-09