Amazon Bedrock fine-tuning: the 2026 guide.
How to customize models on Amazon Bedrock: when to fine-tune vs use prompting or RAG, supervised vs reinforcement fine-tuning vs distillation, which models support it, preparing JSONL data, running a job, hosting the result, evaluating it, and what it costs.
- Bedrock fine-tuning trains a supported base model (Amazon Nova, Meta Llama 3.x, Claude 3 Haiku and others) on your examples so it learns a style, format or narrow skill. Bedrock offers supervised fine-tuning, reinforcement fine-tuning and distillation. Try prompting and RAG first: they change behavior or add facts with no training cost.
- Training is billed by tokens processed (training tokens × epochs), plus a monthly storage charge per custom model. To serve the model you buy Provisioned Throughput, or, for some base models (Nova Lite, Nova 2 Lite, Nova Micro, Nova Pro, Llama 3.3 70B), use an on-demand deployment that bills for use.
- Fine-tuning pays off for narrow, stable, high-volume tasks where prompting cannot make the behavior reliable. Evaluate the tuned model head-to-head against the base model before committing to hosting.
IWhat fine-tuning on Amazon Bedrock is
Fine-tuning continues training a foundation model on your own data so it does a specific task better. On Bedrock the process is managed: you supply data, AWS runs the training, and you get a private custom model in your account.
A base model knows language in general but not your output format, tone or narrow task. Fine-tuning shows it examples of the input it will see and the output you want, and adjusts its weights toward that output. The result is a custom model derived from the base.
On Bedrock you do not provision GPUs or write a training loop. You put data in S3, start a customization job on a supported base model, set a few hyperparameters, and Bedrock does the rest. Your training data and custom model stay in your account.
Fine-tuning is not RAG. RAG gives a model new facts at request time by retrieving documents into the prompt. Fine-tuning changes behavior and is poor at teaching facts that change. See the Bedrock model customization guide.
IIThe ways to customize a model on Bedrock
Picking the wrong method is the most common and most expensive mistake. Here is the menu, lightest first.
Continued pre-training on unlabeled domain text is not among the methods AWS currently lists for Bedrock. If you need it, train on SageMaker and bring the result in with Custom Model Import, if the architecture is supported.
1. Prompt engineering
Better instructions, few-shot examples, a structured system prompt and constrained output format. No training and no custom model. Modern models plus careful prompting solve many "we need to fine-tune" requests. Prompt Management and prompt caching help here.
2. RAG (Knowledge Bases)
Retrieve chunks from your documents into the prompt so the model answers from your knowledge. Use it whenever the problem is "the model does not know our content." Bedrock Knowledge Bases manage ingestion, embeddings, the vector store and retrieval. See RAG on AWS.
3. Supervised fine-tuning (SFT)
Train on labeled prompt/response pairs to lock in a behavior: a strict JSON shape, a brand voice, a niche classification. Use it when you have good labeled examples and prompting is not reliable enough.
4. Reinforcement fine-tuning (RFT)
Instead of labeled pairs, you define reward functions (for example in AWS Lambda) that score responses, and the model learns from that feedback. AWS positions it for tasks with measurable success criteria: code, math, structured analysis, instruction following, or cases where labeled examples are hard to collect. See the reinforcement fine-tuning guide.
5. Distillation
A large "teacher" model generates responses to your prompts, and Bedrock fine-tunes a smaller "student" model on them. You get near-teacher quality on a narrow task at the student's cost and speed. Good for high-volume, narrow workloads.
IIIWhich customization method should you use?
Match the method to what is actually wrong with the base model's output.
- The model does not know our facts → RAG: Fine-tuning is bad at facts and goes stale when data changes. Use Knowledge Bases.
- The prompt or examples are not good enough → prompt engineering: Free and instant. Most fine-tuning requests end here.
- We need a locked-in format, style or narrow skill and have labeled data → supervised fine-tuning: The classic fine-tuning use case.
- We can score answers but cannot easily label them → reinforcement fine-tuning: Define a reward function and let the model learn from it.
- We need big-model quality at small-model cost on a narrow task → distillation: Teach a small model to mimic a large one.
- Combine them: A common stack is RAG for facts, a light fine-tune for output format, and prompt caching for cost.
IVThe methods compared
What each method changes, what data it needs, and whether it leaves you with a hosting decision.
| Method | What it changes | Data needed | Teaches new facts? | Hosting |
|---|---|---|---|---|
| Prompt engineering | Behavior, via instructions | None, or a few examples | Only what you paste in | None (base model) |
| RAG | Facts available at request time | Your documents | Yes | None (base model plus retrieval) |
| Supervised fine-tuning | Style, format, narrow skill | Labeled pairs (JSONL) | Poorly | Provisioned Throughput or on-demand (some models) |
| Reinforcement fine-tuning | Behavior, via rewards | Prompts plus reward functions | No | Custom model hosting |
| Distillation | Small model mimics a big one | Prompts (teacher writes answers) | Inherits teacher behavior | Custom model hosting |
VWhich models support fine-tuning on Bedrock
Not every Bedrock model is tunable, and the list changes. This is the list AWS showed on 2026-10-10.
Newer frontier models are often not tunable on Bedrock. In that case, get the behavior through prompting and RAG, or distill into a supported student. The base model you pick also affects hosting cost: a small model is cheaper to serve. See the fine-tuning guide.
| Provider | Models | Region |
|---|---|---|
| Amazon | Nova 2 Lite, Nova Lite, Nova Micro, Nova Pro, Nova Canvas | us-east-1 |
| Amazon | Titan Image Generator G1 v2, Titan Multimodal Embeddings G1 | us-east-1, us-west-2 |
| Anthropic | Claude 3 Haiku | us-west-2 |
| Meta | Llama 3.1 8B and 70B Instruct; Llama 3.2 1B, 3B, 11B and 90B Instruct; Llama 3.3 70B Instruct | us-west-2 |
VITraining data: format and preparation
Fine-tuning quality is mostly a data problem. Bedrock expects JSONL in S3, one example per line.
Format. JSONL: one self-contained JSON object per line. For supervised fine-tuning, each line pairs an input with the exact output you want; chat-formatted models use a messages-style schema. Field names depend on the model, so check its schema in the docs. Upload training data (and ideally a separate validation file) to S3 in the job's Region.
Mirror production. Training prompts should look like real prompts, including any system instruction and retrieved context, and outputs should match the exact format, length and tone you want. A few hundred to a few thousand clean, representative examples usually beat a larger noisy set.
Hygiene. Remove duplicates and contradictions, balance the cases you care about, hold out a validation set, and strip PII and secrets you do not want in a model.
- One JSON object per line, UTF-8.
- The schema your chosen model expects.
- Prompts that look like real production prompts.
- A separate validation file.
- No duplicates, contradictions, PII or secrets.
- Files in S3 in the same Region as the job.
VIIRunning a fine-tuning job, and hosting the result
Running the job is simple. The hosting choice afterwards decides whether fine-tuning makes economic sense.
Running the job
Create a model customization job in the console or API: pick the base model, point at training and validation data in S3, name the output, set hyperparameters (epochs, learning-rate multiplier, batch size), and give it an IAM role. Bedrock runs the training and reports training and validation loss. The custom model appears in your account.
Hosting: two options
Provisioned Throughput. You buy dedicated model units and pay an hourly rate while the model is deployed, whether or not you send traffic. A longer commitment term lowers the rate. This works for any custom model and suits high, steady volume.
On-demand deployment. For some base models AWS lets you deploy a custom model for on-demand inference and pay for what you use (see deploy a custom model for on-demand inference). On 2026-10-10 AWS listed Amazon Nova Lite, Nova 2 Lite, Nova Micro and Nova Pro (us-east-1) and Meta Llama 3.3 70B Instruct (us-west-2). This changes the economics for low or spiky traffic.
The training charge is usually small next to Provisioned Throughput hosting. If your model qualifies for on-demand deployment, low-traffic fine-tunes become far more practical. See Provisioned Throughput.
VIIIEvaluating the tuned model, and when it is worth it
A finished fine-tune is a hypothesis. Prove it beats the base model before you pay to host it.
- Read the loss curves: Validation loss falling with training loss is healthy. Validation loss rising while training loss falls means overfitting; cut epochs or improve data.
- Evaluate on a held-out set: Run the same unseen prompts through the base and custom models and score them with a task metric (schema validity, accuracy, a rubric or model-as-judge). Bedrock model evaluation can systematize this.
- Compare against the cheap options: The bar is beating the base model with good prompting and RAG, not beating nothing.
- Count the full cost: Training, monthly model storage, hosting, and the effort to maintain the dataset as the task changes.
Fine-tuning wins when the task is narrow, stable, high-volume and format-sensitive. It loses when traffic is low and no on-demand deployment is available, the task keeps changing, or the real gap was missing facts (RAG) or a weak prompt.
IXWho can fund a fine-tuning project
Training, model storage, hosting and the data pipeline around them all draw down AWS credits automatically.
Partner-filed funding: a vetted AWS partner can file POC (proof of concept) funding to test whether a custom model beats the base model, or incremental workload (IW) funding if you already run on AWS and this is a new AI feature. For startups this is typically $50K to $150K. Depends on the workload; AWS approves.
AWS Activate: you apply yourself. Activate Founders gives self-funded startups $1,000 to start, up to $5,000. Activate Portfolio gives up to $200,000 with the Org ID from your VC, accelerator or angel group, if it is an Activate Provider. A consulting partner cannot file Activate for you.
CloudRoute matches you with one vetted AWS ML partner who can prepare the dataset, run the job, set up evaluation, tell you honestly whether to fine-tune at all, and check which funding fits. Startups pay CloudRoute nothing. Partners pay CloudRoute a monthly membership plus 10% of AWS-approved funding.
| Funding | Who files it | Amount | When it fits |
|---|---|---|---|
| Migration Acceleration Program (MAP) | Your matched AWS partner | $50K to $150K for startups. Depends on the workload; AWS approves. | You are not on AWS yet and want to migrate. |
| Proof of concept (POC) funding | Your matched AWS partner | $50K to $150K for startups. Depends on the workload; AWS approves. | You want to test something, whether or not you are on AWS today. |
| Incremental workload (IW) funding | Your matched AWS partner | $50K to $150K for startups. Depends on the workload; AWS approves. | You are already on AWS and are launching new work: a new AI feature, a new customer or a new contract. |
| AWS Activate Founders | Founder applies | $1,000 to start, up to $5,000 | Self-funded startups with no investor or accelerator behind them. |
| AWS Activate Portfolio | Founder applies | up to $200,000 | Startups backed by a VC firm, accelerator or angel investor that is an AWS Activate Provider. |
| AWS Credits for AI Startups | Invite only (AWS account manager) | $200,000 or more, by invitation | AI startups ready to scale after Activate Portfolio, invited by AWS. |
Fine-tuning vs RAG vs prompt engineering vs distillation
Match the row to the problem you actually have.
| Method | Best when | Effort | Adds facts? | Changes behavior? | Reach for it |
|---|---|---|---|---|---|
| Prompt engineering | The prompt or examples are weak | Lowest | Only inline | Yes | First, always |
| RAG | The model lacks your facts or latest data | Low to medium | Yes | No | For any knowledge gap |
| Supervised fine-tuning | You need a locked-in format or skill and have labeled data | Medium | No | Yes | With good labeled data |
| Reinforcement fine-tuning | You can score answers but not easily label them | Medium to high | No | Yes | Measurable tasks |
| Distillation | You need big-model quality at small-model cost | Medium | Inherits teacher | Yes | High-volume narrow tasks |
How a match works (example)
- You tell us about the work
You fill in a short form: what you run today, what you want to move, test or build, your stage and your region.
- We pick one partner
We read your answers and pick one vetted AWS consulting partner that fits the work, the stack and the region. We reply within 24 hours; your matched partner reaches out within 5 business days.
- The partner checks POC funding
Your matched partner scopes the test with you. Then it checks whether the test qualifies for POC (proof of concept) funding. A POC can fit whether or not you run on AWS today.
- The partner files, AWS decides
If the work qualifies, the partner files the funding request with AWS. AWS reviews it and decides. The amount depends on the workload.
- You pay nothing
CloudRoute is free for startups. Partners pay CloudRoute a monthly membership plus 10% of AWS-approved funding.
AWS Activate credits are separate, and you apply for them yourself. Use Activate Founders if no investor backs you. Use Activate Portfolio with the Org ID or invite from your VC, accelerator, incubator or angel group if it is an Activate Provider. A consulting partner cannot apply for Activate on your behalf.
Common questions
What is fine-tuning on Amazon Bedrock?
When should I fine-tune versus use RAG or prompt engineering?
Which models can I fine-tune on Bedrock?
What is reinforcement fine-tuning on Bedrock?
Does a fine-tuned model need Provisioned Throughput?
How much does fine-tuning on Bedrock cost?
Is fine-tuning worth it?
Can AWS funding cover a fine-tuning project?
Fine-tune with a partner who can file the funding
A vetted AWS ML partner can prepare the data, run the job, evaluate the result and tell you honestly whether to fine-tune at all. The same partner checks whether the work qualifies for POC or IW funding: typically $50K to $150K for startups. Depends on the workload; AWS approves. CloudRoute is free for startups.