Fine-tuning - verified
Amazon Bedrock fine-tuning: status, methods and prices
Offered. 3 offerings recorded: SFT, RFT and Distillation. Training is priced per 1M training tokens, per training hour, per training image, fixed fees: per job or per month and billed at another rate, in Amazon Bedrock's own units. Weight export: Not stated.
Recorded as: Amazon Bedrock.
Read from docs.aws.amazon.com, aws.amazon.com. Compare every fine-tuning API
- Status
- Offered
- Offerings
- 3
- Methods
- SFT, RFT, Distillation
- Export weights
- Not stated
Status
Is Amazon Bedrock fine-tuning still available?
Offered
Distillation not available for Anthropic models, no restore timeline
“Distillation is not currently available for Anthropic models on Amazon Bedrock. There is no confirmed timeline for when Anthropic distillation will be restored.”
All 3 offerings
Supervised fine-tuning
- Status
GA (inferred)
Inferred: the page does not state this outright.
“With Amazon Bedrock, you can train a foundation model to improve performance on specific tasks”
docs.aws.amazon.com - Tunable base models
amazon.nova-2-lite-v1:0:256kamazon.nova-canvas-v1:0amazon.nova-lite-v1:0:300kamazon.nova-micro-v1:0:128kamazon.nova-pro-v1:0:300kamazon.titan-image-generator-v2:0amazon.titan-embed-image-v1:0anthropic.claude-3-haiku-20240307-v1:0:200kmeta.llama3-1-70b-instruct-v1:0:128kmeta.llama3-1-8b-instruct-v1:0:128kmeta.llama3-2-11b-instruct-v1:0:128kmeta.llama3-2-1b-instruct-v1:0:128kmeta.llama3-2-3b-instruct-v1:0:128kmeta.llama3-2-90b-instruct-v1:0:128kmeta.llama3-3-70b-instruct-v1:0:128k
Full list from the same table; four representative rows quoted.
“Amazon · Nova 2 Lite · amazon.nova-2-lite-v1:0:256k · us-east-1”
docs.aws.amazon.com “Amazon · Nova Pro · amazon.nova-pro-v1:0:300k · us-east-1”
docs.aws.amazon.com “Anthropic · Claude 3 Haiku · anthropic.claude-3-haiku-20240307-v1:0:200k · us-west-2”
docs.aws.amazon.com “Meta · Llama 3.3 70B Instruct · meta.llama3-3-70b-instruct-v1:0:128k · us-west-2”
docs.aws.amazon.com - Training price
Supervised fine-tuning training prices in Amazon Bedrock's own units Model Price Unit Quote Llama 3.2 Instruct (1B) $0.50 per 1M tokens trained “Llama 3.2 Instruct (1B) · $0.50 · $1.95 · $23.00”aws.amazon.com Llama 3.2 Instruct (3B) $1.10 per 1M tokens trained “Llama 3.2 Instruct (3B) · $1.10 · $1.95 · $23.00”aws.amazon.com Llama 3.2 Instruct (11B) $3.50 per 1M tokens trained “Llama 3.2 Instruct (11B) · $3.50 · $1.95 · $23.00”aws.amazon.com Llama 3.2 Instruct (90B) $7.90 per 1M tokens trained “Llama 3.2 Instruct (90B) · $7.90 · $1.95 · $30.00”aws.amazon.com Llama 3.1 Instruct (8B) $1.49 per 1M tokens trained “Llama 3.1 Instruct (8B) · $1.49 · $1.95 · $24.00”aws.amazon.com Llama 3.1 Instruct (70B) $7.99 per 1M tokens trained “Llama 3.1 Instruct (70B) · $7.99 · $1.95 · $24.00”aws.amazon.com Amazon Nova Canvas $0.005 per image seen “Amazon Nova Canvas · $0.005 · $1.95 · $60.50 · $55.00 · $30.25”aws.amazon.com storage (all custom models above) $1.95 per custom model per month “Llama 3.1 Instruct (8B) · $1.49 · $1.95 · $24.00”aws.amazon.com - Tuned-model inference
Llama 3.2 Instruct (1B/3B/11B): $23.00 (per model unit per hour (no-commit Provisioned Throughput))
“Llama 3.2 Instruct (1B) · $0.50 · $1.95 · $23.00”
aws.amazon.com Llama 3.1 Instruct (8B/70B): $24.00 (per model unit per hour (no-commit Provisioned Throughput))
“Llama 3.1 Instruct (70B) · $7.99 · $1.95 · $24.00”
aws.amazon.com Amazon Nova Canvas: $60.50 no commit / $55.00 1-month / $30.25 6-month (per model unit per hour)
“Amazon Nova Canvas · $0.005 · $1.95 · $60.50 · $55.00 · $30.25”
aws.amazon.com
- Hosting requirement
Provisioned Throughput, or a custom model deployment for on-demand inference (model-dependent)
“Purchase Provisioned Throughput – Purchase”
docs.aws.amazon.com “Deploy custom model for on-demand inference”
docs.aws.amazon.com “With on-demand inference, you only pay for what you use and you don't need to set up”
docs.aws.amazon.com - Export weights
Not stated
- Data limits
- Not stated
- Notes
Claude 3 Haiku is the only Anthropic model in the fine-tuning table (us-west-2); its training price was not in the fetched pricing text. Anthropic's own cookbook says a fine-tuned Claude model on Bedrock must be hosted with Provisioned Throughput. Nova text-model (Nova 2 Lite, Micro, Lite, Pro) SFT prices: null for the same rendering reason. Llama 3.3 70B is tunable but no price row was visible.
Reinforcement fine-tuning
- Status
- Not stated
- Tunable base models
amazon.nova-2-lite-v1:0:256kopenai.gpt-oss-20bqwen.qwen3-32b
“Amazon · Nova 2 Lite · amazon.nova-2-lite-v1:0:256k · US East (N. Virginia) · us-east-1”
docs.aws.amazon.com “OpenAI · gpt-oss-20B · openai.gpt-oss-20b · US West (Oregon) · us-west-2”
docs.aws.amazon.com “Qwen · Qwen3 32B · qwen.qwen3-32b · US West (Oregon) · us-west-2”
docs.aws.amazon.com - Training price
Reinforcement fine-tuning training prices in Amazon Bedrock's own units Model Price Unit Quote gpt-oss-20b $80.00 per training hour “gpt-oss-20b · $ 80.00 · $ 0.09 · $ 0.39 · $ 1.95”aws.amazon.com Qwen3 32B $80.00 per training hour “Qwen3 32B · $ 80.00 · $ 0.20 · $ 0.78 · $ 1.95”aws.amazon.com billing basis hourly per hour “The entire training workflow is billed at an hourly rate.”aws.amazon.com - Tuned-model inference
gpt-oss-20b (RFT): $0.09 input / $0.39 output; $1.95 storage per month (per 1M tokens (on-demand))
“gpt-oss-20b · $ 80.00 · $ 0.09 · $ 0.39 · $ 1.95”
aws.amazon.com Qwen3 32B (RFT): $0.20 input / $0.78 output; $1.95 storage per month (per 1M tokens (on-demand))
“Qwen3 32B · $ 80.00 · $ 0.20 · $ 0.78 · $ 1.95”
aws.amazon.com
- Hosting requirement
on-demand inference available immediately after RFT
“After training completes, customers can immediately use the resulting fine tuned model for on-demand inference.”
aws.amazon.com - Export weights
Not stated
- Data limits
- Not stated
- Your training data
Data stays in AWS during customization
“Your proprietary data never leaves AWS's secure, governed environment during the customization process.”
docs.aws.amazon.com - Notes
Nova 2 Lite RFT hourly price was not in the fetched pricing text. Rewards are defined with AWS Lambda or model-as-a-judge graders; training uses GRPO.
Model distillation
- Status
- Not stated
- Tunable base models
amazon.nova-lite-v1:0:300kamazon.nova-micro-v1:0:128kamazon.nova-pro-v1:0:300kmeta.llama3-1-8b-instruct-v1:0:128kmeta.llama3-1-70b-instruct-v1:0:128kmeta.llama3-2-1b-instruct-v1:0:128kmeta.llama3-2-3b-instruct-v1:0:128kmeta.llama3-3-70b-instruct-v1:0:128k
Student models. Teachers: Nova Pro, Nova Premier, Llama 3.1 405B, Llama 3.1 70B, Llama 3.3 70B.
“Amazon · Nova Pro · amazon.nova-pro-v1:0 · Both · Nova Lite, Nova Micro”
docs.aws.amazon.com “Meta · Llama 3.1 405B · meta.llama3-1-405b-instruct-v1:0 · On demand · Llama 3.1 8B, Llama 3.1 70B, Llama 3.2 1B, Llama 3.3 70B”
docs.aws.amazon.com - Training price
Model distillation training prices in Amazon Bedrock's own units Model Price Unit Quote teacher synthesis calls teacher model on-demand inference rates per teacher inference call “These charges will be billed at the on-demand inference rates of the teacher model.”docs.aws.amazon.com - Tuned-model inference
- Not stated
- Hosting requirement
Provisioned Throughput (or copy model to another Region and buy PT) for Claude/Llama distillations
“You can either buy provisioned throughput in”
docs.aws.amazon.com - Export weights
Not stated
- Data limits
synthesis may expand the dataset up to 15k prompt-response pairs
“may increase the size of the fine-tuning dataset to a maximum of 15k prompt-response pairs”
docs.aws.amazon.com
- Your training data
Customer data not used to train other models for public use
“Amazon Bedrock doesn't use your data to train any other teacher or student model for public use.”
docs.aws.amazon.com - Notes
Student fine-tuning cost itself is not separately stated on the distillation page; the general custom-models note charges per training token plus storage.
Research notes
Bedrock's customization overview now lists three methods: supervised fine-tuning, reinforcement fine-tuning, and distillation. Continued pre-training is no longer listed on that page (the pricing page still carries a 'Customization (fine-tuning and continued pretraining) pricing' example for Titan). Nova text-model and Claude 3 Haiku customization prices are rendered client-side by a region selector and did not appear in the fetched page text, so they are null. Training is charged per token processed (tokens x epochs) plus monthly storage per custom model.
Plain language
The methods, explained
- SFTSupervised fine-tuning
- You show the model example prompts with the answers you want, and it learns to answer like that. The default method almost everywhere.
- DPODirect preference optimization (preference tuning)
- You give pairs of answers to the same prompt, one preferred and one not, and the model learns to lean toward the preferred kind. Good for tone and style.
- RFTReinforcement fine-tuning
- The model attempts tasks and a grader (your code, or another model) scores each attempt; training pushes it toward higher-scoring answers. Often billed by training hour, plus the grader's tokens.
- Vision SFTVision fine-tuning
- Supervised fine-tuning where the examples include images, so the model gets better at a visual task.
- DistillationDistillation
- A large teacher model writes answers that are used to fine-tune a smaller, cheaper student model. You pay for the teacher's calls as well as the training.
- LoRALoRA (adapter) fine-tuning
- Trains a small set of extra weights, an adapter, on top of the frozen base model. Cheaper and faster than full fine-tuning; the adapter is served alongside the base model.
- FullFull-parameter fine-tuning
- Updates every weight in the model. Costs more than LoRA and produces a complete copy of the model.
- CPTContinued pre-training
- Keeps training the model on a large body of raw, unlabeled text from your domain, so it absorbs vocabulary and facts before any instruction tuning.
FAQ
Does Amazon Bedrock offer fine-tuning?
Offered. 3 offerings: SFT, RFT and Distillation.
How much does Amazon Bedrock fine-tuning cost?
Per 1M training tokens: Llama 3.2 Instruct (1B) $0.50, Llama 3.2 Instruct (3B) $1.10, Llama 3.2 Instruct (11B) $3.50, Llama 3.2 Instruct (90B) $7.90 and 2 more. Per training hour: gpt-oss-20b $80.00, Qwen3 32B $80.00. Per training image: Amazon Nova Canvas $0.005. Fixed fees: per job or per month: storage (all custom models above) $1.95. Billed at another rate: billing basis hourly, teacher synthesis calls teacher model on-demand inference rates. Prices are in Amazon Bedrock's own units and are not converted.
Can you download a model you fine-tune on Amazon Bedrock?
Not stated. The pages we checked do not say whether tuned weights can be exported.
Sources
- https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html
- https://aws.amazon.com/bedrock/pricing/
- https://docs.aws.amazon.com/bedrock/latest/userguide/prequisites-model-distillation.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/custom-model-fine-tuning.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/model-customization-use.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/reinforcement-fine-tuning.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/model-distillation.html
Last verified . This is a directory of what each provider's own pages say, not a quality ranking. Prices are copied in the provider's own unit and never converted, so rows in different units are not comparable.
Newsletter
Know when a provider changes its fine-tuning terms
One email per week. New model releases, deprecations and pricing changes, and what they mean for what you build.