When you’re operating Amazon Nova 1 fashions on Amazon Bedrock, you is likely to be seeking to develop your context window dimension, deepen reasoning capabilities, or combine exterior instruments for net search and code execution. Amazon Nova 2 fashions deal with these constraints whereas bettering efficiency on reasoning, agentic AI, and power use benchmarks.
Amazon Nova 2 Lite achieves increased scores throughout downside identification, resolution completeness, and logical coherence benchmarks, whereas sustaining quick response occasions for high-volume workloads. When you’re utilizing Nova 1 Lite for buyer help automation, doc processing, or agentic AI functions, you’ll doubtless see measurable features in accuracy and throughput upon migrating to Nova 2 Lite. This publish focuses on migration to Nova 2 Lite, which is mostly out there and prepared for manufacturing use. Nova 2 expands the context window to at least one million tokens, which helps richer in-context studying and permits for processing longer paperwork in a single request. You additionally acquire entry to prolonged considering, built-in net grounding, and a code interpreter. These options can improve your present functions with minimal code modifications.
On this publish, you’ll discover ways to migrate from Nova 1 to Nova 2 on Amazon Bedrock. We cowl mannequin mapping, API modifications, code examples utilizing the Converse API, steerage on configuring new capabilities, and a abstract of use instances. We conclude with a migration guidelines that will help you plan and execute your transition.
Migration paths
The next are really useful migration paths for the Nova 1 fashions. For Professional and Premier migrations, consider with prolonged considering enabled to confirm high quality in your workloads.
- When you’re operating Amazon Nova 1 Lite, the migration path is easy: Nova 2 Lite is a direct improve that maintains the identical enter modalities (textual content, picture, and video) whereas including prolonged considering, built-in instruments, and a context window that expands from 300K to 1M tokens.
- When you’re operating Amazon Nova 1 Professional, we advocate upgrading to Nova 2 Lite. Whereas this may seem to be a tier change, Nova 2 Lite delivers improved reasoning capabilities at aggressive price-performance. The prolonged considering characteristic in Nova 2 Lite, mixed with its 1M token context window, permits it to deal with workloads that beforehand required Nova 1 Professional’s bigger mannequin dimension.
- When you’re operating Amazon Nova Premier, take into account migrating to Nova 2 Lite, particularly for agentic and power use workloads. As documented within the Amazon Nova 2 Technical Report, Nova 2 Lite surpasses Premier in multi-step problem-solving at 7x decrease price and as much as 5x sooner inference. As a result of Nova 2 Lite with prolonged considering enabled remains to be cheaper and sooner than Premier, we advocate evaluating throughout reasoning effort ranges to confirm that it meets your high quality necessities.
Use instances and buyer tales
Since Nova 2 Lite launched at re:Invent 2025, you should utilize the mannequin for a wide range of use instances. Nova 2 Lite excels at driving excessive throughput functions that require a mixture of value, efficiency, and pace. Widespread use instances that prospects have deployed Nova 2 Lite in manufacturing embody:
Pure Language Processing (NLP) duties
Nova 2 Lite is used for all kinds of NLP duties, from easy textual content predictors to logic-based suggestions. The Nova 2 fashions excel at summarization, classification, and search based mostly use instances. Nova 2 Lite scores 80.9% on MMLU Professional and 70.8% on IF-Bench (with each increased than some comparable fashions). Nova 2 Lite can be designed for low latency, making it appropriate for functions requiring fast responses, as is typical with most NLP duties.
Siemens international search runs on Nova 2 Lite, offering a 300% enchancment in search pace and 70% price reductions in comparison with their prior giant language mannequin (LLM) resolution. The total case research will be discovered on the AWS case research web page.
Clever Doc Processing (IDP)
Nova 2 fashions with reasoning improve IDP by shifting past fundamental textual content extraction to true semantic understanding of paperwork. They’ll course of unstructured and semistructured content material similar to contracts, invoices, varieties, and experiences, and decide what data means in context, not solely the place it seems. Reasoning permits these fashions to hyperlink associated data throughout sections, pages, and tables, deal with variations in format and language, and infer lacking or implicit particulars, making doc processing extra sturdy and scalable. Reasoning provides worth by (1) understanding context. For instance, distinguishing a contract’s begin date from an finish date even when each seem as related timestamps. (2) linking and inferring throughout pages. For instance, tying bill line gadgets to totals/taxes and checking the mathematics (even filling small gaps when textual content is unreadable). (3) normalizing variation—treating “quantity payable,” “steadiness due,” and “whole excellent” as the identical idea. (4) enabling smarter actions like validation, anomaly detection, and Q&A—flagging mismatched tax calculations or figuring out auto-renewal clauses from associated sections. Reasoning turns IDP into a better, extra human-like system that understands and validates paperwork, enabling extra correct automation, much less handbook overview, and sooner adaptation to new ontology.
Multi-step agentic workflows
Nova 2 Lite reveals appreciable enchancment over the Nova 1 fashions in software calling capabilities, particularly if a number of instruments are a part of the workflow. Total, the Nova 2 household excels at finishing complicated, multi-step agentic workflows with refined software use and planning capabilities. Nova 2 Lite scores increased than comparable fashions in τ2-bench Telecom, at 76.0%. We advocate testing multi-step agentic use instances with reasoning on, beginning with reasoning set to “Low” after which working upwards, if required.
Nova 2 Lite has empowered Trellix to proceed to automate its safety alert triage duties and higher empower safety groups, within the areas of menace detection, evaluation, and classification. With the discharge of the brand new mannequin, Trellix noticed extra dependable software calling capabilities with no failures, a 39% accuracy increase in menace classification, and three.4x extra detailed responses with technical evaluation. Watch Trellix’s CTO clarify the answer on this video.
AWS Rework is a multi-agent, multiuser system that makes use of dozens of instruments to modernize complicated infrastructure methods and code bases. AWS Rework breaks down every consumer request into plenty of agent and power calls which can be routed to a combination of fashions to steadiness price, pace, and accuracy. Evaluations confirmed that Nova 2 Lite improves software calling effectivity for code modernization in AWS Rework by as much as 60%.
What’s modified
The Nova 2 household introduces capabilities that develop how one can construct AI functions, notably for agentic AI workloads.
Prolonged considering with developer controls
Nova 2 Lite can motive by means of complicated issues earlier than responding. You management the depth of reasoning by means of the reasoningConfig parameter with low, medium, or excessive effort ranges.
Native software use and built-in net grounding and code interpreter
Nova 2 Lite is designed for software use, together with help for MCP servers and parallel software chaining. It additionally consists of built-in net grounding for real-time data with citations and a code interpreter for executing Python code straight.
Prolonged context for large-scale inputs
The context window expands from 300K to 1M tokens, and most output tokens improve from 10K to 65K. With these will increase, you’ll be able to course of bigger paperwork, code bases, and longer multi-turn workflows in a single request.
The next desk summarizes the technical specs throughout Nova 1 fashions and Nova 2 Lite:
Specification
Nova 1 Lite
Nova 1 Professional
Nova 2 Lite
Context Window
300K
300K
1M
Max Output Tokens
10K
10K
65K
Enter Modalities
Textual content, picture, video
Textual content, picture, video
Textual content, picture, video
Prolonged Pondering
No
No
Sure
Constructed-in Instruments
No
No
Sure
Customization
Sure (incl. Nova Forge)
Sure (incl. Nova Forge)
Sure (incl. Nova Forge)
Efficiency and pricing
Unbiased benchmarks from Synthetic Evaluation present that Nova 2 Lite scores above common on intelligence amongst comparable fashions and is notably quick, making it well-suited for high-volume workloads the place throughput issues. On agentic-specific benchmarks, Nova 2 Lite performs nicely on Software Use (Tau2 Telecom). Sturdy instruction following is important for agentic workloads the place the mannequin should reliably execute multi-step plans. For detailed benchmark methodology and extra outcomes, see the Amazon Nova 2 Technical Report.
Pricing for the Nova 2 household displays the expanded capabilities whereas remaining aggressive inside their respective tiers.
Mannequin
Enter (per 1M tokens)
Output (per 1M tokens)
Nova 1 Lite
$0.06
$0.24
Nova 1 Professional
$0.80
$3.20
Nova 2 Lite
$0.30
$2.50
For Nova 2 Lite, pricing is aggressive towards related fashions in its class. For the complete announcement and extra particulars, see the Amazon Nova 2 launch publish.
API modifications and code updates
Replace your mannequin ID to reference Nova 2 Lite, which is accessible by means of International Cross-Area Inference System (CRIS), US CRIS, EU CRIS, and JP CRIS endpoints. When accessing fashions from a US area, use the us. prefix (for instance, us.amazon.nova-2-lite-v1:0); when accessing from exterior the US, use the worldwide. prefix (for instance, international.amazon.nova-2-lite-v1:0). The underlying API construction stays constant, so present integrations will work with minimal modifications past the mannequin ID and any new options that you simply select to allow.
Breaking change: When maxReasoningEffort is ready to excessive, you can not use maxTokens, temperature, topP, or topK. Requests that embody these parameters with excessive effort reasoning will return an error. To resolve this, take away the inferenceConfig block completely out of your request.
New options and unchanged behaviors
The next desk summarizes new options and unchanged behaviors between Nova 1 and Nova 2:
Class
Change
Particulars
New
Prolonged considering
Allow reasoning utilizing reasoningConfig in additionalModelRequestFields. Helps three effort ranges: low, medium, and excessive.
New
Reasoning content material in responses
Responses embody a reasoningContent area when reasoning is enabled. Content material shows as [REDACTED] however reasoning tokens are billed.
New
Constructed-in instruments
Internet grounding and code interpreter can be found with out exterior integrations.
New
Elevated output capability
Most output tokens elevated from 10K to 65,536.
New
Expanded context window
Nova 2 Lite helps 1M tokens (up from 300K in Nova 1 Lite).
Unchanged
topK parameter location
topK remains to be handed by means of additionalModelRequestFields, not inferenceConfig.
Unchanged
Doc help
Doc enter remains to be Converse API solely (not supported in Invoke API).
Unchanged
Timeout configuration
Boto3’s default learn timeout is 60 seconds, however prolonged considering requests can take as much as 60 minutes. Configure read_timeout=3600 in your SDK shopper.
Configuring new capabilities
This part covers the essential migration steps, prolonged considering configuration, and learn how to use built-in instruments like net grounding and code interpreter.
Primary migration – replace the mannequin ID
For functions that don’t require prolonged considering, the migration is easy. Replace the mannequin ID and your present code will work:
import boto3
bedrock = boto3.shopper(‘bedrock-runtime’, region_name=”us-east-1″)
# Nova 1
# modelId=’us.amazon.nova-lite-v1:0′
# Nova 2
response = bedrock.converse(
modelId=’us.amazon.nova-2-lite-v1:0′,
system=[{‘text’: ‘You are a helpful assistant’}],
messages=[
{
‘role’: ‘user’,
‘content’: [{‘text’: ‘Explain cloud computing in simple terms.’}]
}
],
inferenceConfig={
‘maxTokens’: 1024,
‘temperature’: 0.7,
‘topP’: 0.9
}
)
print(response[‘output’][‘message’][‘content’][0][‘text’])
Prolonged considering
Prolonged considering permits Nova 2 mannequin to motive by means of complicated issues earlier than producing responses. This functionality is disabled by default to optimize for pace and price on easy queries. When your workload requires multi-step reasoning, you’ll be able to allow it by means of the reasoningConfig parameter.
Configuring reasoning effort ranges
Nova 2 Lite offers three effort ranges that management how a lot reasoning the mannequin performs:
- Low effort (maxReasoningEffort: “low”) is really useful for duties with added complexity that profit from structured considering. Use low effort for code overview and enchancment options, evaluation duties that require consideration of a number of elements, or problem-solving situations that profit from a methodical method. Low effort improves accuracy on compound duties with out requiring deep multi-step planning. We advocate beginning with low effort and dealing your manner as much as increased effort ranges as wanted on your particular use case.
- Medium effort (maxReasoningEffort: “medium”) is really useful for multi-step duties and coding workflows. Use medium effort for software program growth and debugging the place the mannequin should perceive present code construction, code technology that requires coordination throughout a number of information, multi-step calculations with interdependencies, or planning duties with a number of constraints. Medium effort is really useful for agentic workflows that coordinate a number of instruments and require the mannequin to take care of context throughout sequential operations.
- Excessive effort (maxReasoningEffort: “excessive”) is really useful for STEM reasoning and superior problem-solving. Use excessive effort for superior mathematical issues and proofs, scientific evaluation, complicated system design, or important decision-making situations. When utilizing excessive effort, don’t embody maxTokens, temperature, topP, or topK.
Instance – Medium Effort Reasoning:
import boto3
bedrock = boto3.shopper(‘bedrock-runtime’, region_name=”us-east-1″)
response = bedrock.converse(
modelId=’us.amazon.nova-2-lite-v1:0′,
system=[{‘text’: ‘You are a helpful assistant’}],
messages=[
{
‘role’: ‘user’,
‘content’: [{‘text’: ‘A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?’}]
}
],
inferenceConfig={
‘maxTokens’: 10000,
‘temperature’: 0.7,
‘topP’: 0.9
},
additionalModelRequestFields={
‘reasoningConfig’: {
‘kind’: ‘enabled’,
‘maxReasoningEffort’: ‘medium’
}
}
)
# Extract the response textual content
content_list = response[‘output’][‘message’][‘content’]
textual content = subsequent((merchandise[‘text’] for merchandise in content_list if ‘textual content’ in merchandise), None)
print(textual content)
Instance – Excessive effort reasoning (notice the parameter restrictions):
import boto3
bedrock = boto3.shopper(‘bedrock-runtime’, region_name=”us-east-1″)
response = bedrock.converse(
modelId=’us.amazon.nova-2-lite-v1:0′,
system=[{‘text’: ‘You are a mathematical reasoning assistant’}],
messages=[
{
‘role’: ‘user’,
‘content’: [{‘text’: ‘Prove that the square root of 2 is irrational.’}]
}
],
# Don’t embody inferenceConfig in any respect with excessive effort reasoning
# (no maxTokens, temperature, topP, or topK)
additionalModelRequestFields={
‘reasoningConfig’: {
‘kind’: ‘enabled’,
‘maxReasoningEffort’: ‘excessive’
}
}
)
# Extract the response textual content
content_list = response[‘output’][‘message’][‘content’]
textual content = subsequent((merchandise[‘text’] for merchandise in content_list if ‘textual content’ in merchandise), None)
print(textual content)
Understanding the response construction
When reasoning is enabled, the response features a reasoningContent area alongside the textual content response:
{
“output”: {
“message”: {
“function”: “assistant”,
“content material”: [
{
“reasoningContent”: {
“reasoningText”: {
“text”: “[REDACTED]”
}
}
},
{
“textual content”: “Based mostly on my evaluation, right here is the really useful structure…”
}
]
}
},
“stopReason”: “end_turn”
}
The reasoning content material presently shows as [REDACTED]. You’re nonetheless charged for reasoning tokens as a result of they contribute to improved output high quality.
Price, latency, and timeout configuration
Prolonged considering will increase each price and latency as a result of the mannequin generates further reasoning tokens earlier than producing the ultimate response. Allow reasoning when accuracy on complicated duties justifies the extra price, and disable it for easy queries the place pace is the precedence.
import boto3
from botocore.config import Config
bedrock = boto3.shopper(
‘bedrock-runtime’,
region_name=”us-east-1″,
config=Config(
read_timeout=3600 # 60 minutes
)
)
Constructed-in instruments: net grounding and code interpreter
Constructed-in system instruments in Nova 2 prolong mannequin capabilities with out requiring customized implementations. You allow these instruments by means of the toolConfig parameter within the Converse API, and Nova routinely decides when to make use of them based mostly on the context of your immediate.
Internet Grounding
To allow Internet Grounding, specify nova_grounding as a system software in your software configuration:
import boto3
from botocore.config import Config
bedrock = boto3.shopper(
‘bedrock-runtime’,
region_name=”us-east-1″,
config=Config(read_timeout=3600)
)
tool_config = {
‘instruments’: [{
‘systemTool’: {
‘name’: ‘nova_grounding’
}
}]
}
response = bedrock.converse(
modelId=’us.amazon.nova-2-lite-v1:0′,
messages=[{
‘role’: ‘user’,
‘content’: [{‘text’: ‘What are the latest developments in quantum computing?’}]
}],
toolConfig=tool_config
)
# Extract textual content with interleaved citations
output_with_citations=””
content_list = response[‘output’][‘message’][‘content’]
for content material in content_list:
if ‘textual content’ in content material:
output_with_citations += content material[‘text’]
elif ‘citationsContent’ in content material:
citations = content material[‘citationsContent’][‘citations’]
for quotation in citations:
url = quotation[‘location’][‘web’][‘url’]
output_with_citations += f’ [{url}]’
print(output_with_citations)
The response consists of citationsContent objects containing the supply URL and area for every quotation. You’re answerable for retaining and displaying these citations to your finish customers.
Internet Grounding is accessible in US AWS Areas solely and requires both BedrockFullAccess or express bedrock:InvokeTool permissions for the amazon.nova_grounding useful resource. Internet Grounding incurs further prices past commonplace inference pricing. For pricing particulars, see Amazon Bedrock Pricing.
Code Interpreter
Code Interpreter permits Nova to execute Python code in remoted sandbox environments. This software is designed for mathematical computations, logical operations, and iterative algorithms the place exact calculation is required. To allow Code Interpreter, specify nova_code_interpreter as a system software:
import boto3
bedrock = boto3.shopper(‘bedrock-runtime’, region_name=”us-east-1″)
tool_config = {
‘instruments’: [{
‘systemTool’: {
‘name’: ‘nova_code_interpreter’
}
}]
}
response = bedrock.converse(
modelId=’us.amazon.nova-2-lite-v1:0′,
messages=[{
‘role’: ‘user’,
‘content’: [{‘text’: ‘Calculate the compound interest on $10,000 at 4.75% annual rate for 7 years, compounded quarterly.’}]
}],
toolConfig=tool_config,
inferenceConfig={‘maxTokens’: 10000, ‘temperature’: 0}
)
# Course of the response
for block in response[‘output’][‘message’][‘content’]:
if ‘toolUse’ in block:
print(f”Code executed:n{block[‘toolUse’][‘input’][‘snippet’]}n”)
elif ‘toolResult’ in block:
outcome = block[‘toolResult’][‘content’][0][‘json’]
print(f”Output: {outcome[‘stdOut’]}”)
elif ‘textual content’ in block:
print(f”Reply: {block[‘text’]}”)
The interpreter runs code in a sandbox and returns ends in a structured format that features stdOut, stdErr, exitCode, and isError fields.
Code Interpreter is accessible within the US East (N. Virginia), US West (Oregon), and Asia Pacific (Tokyo) Areas. To make sure that your requests are routed to a supported Area, use International CRIS. You should manually add InvokeTool permissions to your Id and Entry Administration (IAM) coverage; the default Amazon Bedrock function doesn’t embody this motion.
For detailed documentation on built-in instruments, see Internet Grounding and Software Use within the Amazon Nova documentation.
Framework-specific examples
The earlier examples use the Converse API straight. When you’re utilizing LangChain or Strands Brokers SDK, the migration follows the identical sample: replace the mannequin ID and optionally allow prolonged considering.
Langchain
The langchain-aws package deal offers a ChatBedrockConverse class that integrates with the Amazon Bedrock Converse API. Set up the package deal with pip set up langchain-aws.
from langchain_aws import ChatBedrockConverse
# Nova 1
# llm = ChatBedrockConverse(mannequin=”us.amazon.nova-lite-v1:0″, …)
# Nova 2
llm = ChatBedrockConverse(
mannequin=”us.amazon.nova-2-lite-v1:0″,
region_name=”us-east-1″,
temperature=0.7,
max_tokens=1024
)
messages = [
(“system”, “You are a helpful assistant.”),
(“human”, “Explain cloud computing in simple terms.”),
]
response = llm.invoke(messages)
print(response.content material)
To allow prolonged considering, cross the reasoningConfig by means of additional_model_request_fields:
llm = ChatBedrockConverse(
mannequin=”us.amazon.nova-2-lite-v1:0″,
region_name=”us-east-1″,
max_tokens=10000,
additional_model_request_fields={
“reasoningConfig”: {
“kind”: “enabled”,
“maxReasoningEffort”: “medium”
}
}
)
messages = [
(“system”, “You are a software architect.”),
(“human”, “Design a scalable microservices architecture for an e-commerce platform.”),
]
response = llm.invoke(messages)
# Extract simply the textual content content material (skip reasoning_content blocks)
for block in response.content material:
if isinstance(block, dict) and block.get(‘kind’) == ‘textual content’:
print(block[‘text’])
elif hasattr(block, ‘kind’) and block.kind == ‘textual content’:
print(block.textual content)
Strands Brokers SDK
For agentic functions, the Strands Brokers SDK connects to the Nova API straight. Set up the required packages with pip set up strands-agents strands-amazon-nova.
import os
from strands import Agent
from strands_amazon_nova import NovaAPIModel
# Nova 1
# mannequin = NovaAPIModel(model_id=”nova-lite-v1″, …)
# Nova 2
mannequin = NovaAPIModel(
api_key=os.environ[“NOVA_API_KEY”],
model_id=”nova-2-lite-v1″,
params={
“max_tokens”: 1024,
“temperature”: 0.7
}
)
agent = Agent(mannequin=mannequin)
response = agent(“Clarify cloud computing in easy phrases.”)
print(response.message)
To allow prolonged considering, add the reasoning_effort parameter:
mannequin = NovaAPIModel(
api_key=os.environ[“NOVA_API_KEY”],
model_id=”nova-2-lite-v1″,
params={
“max_tokens”: 10000,
“reasoning_effort”: “medium”
}
)
agent = Agent(mannequin=mannequin)
response = agent(“Design a scalable microservices structure for an e-commerce platform.”)
print(response.message)
Analysis and rollout technique
Earlier than deploying Nova 2 to manufacturing, confirm that the brand new mannequin performs in addition to or higher than Nova 1 in your particular duties. Create a curated dataset of prompts and anticipated outputs that signify your precise manufacturing visitors. For analysis, mix a number of approaches: deterministic metrics like latency, token utilization, and exact-match accuracy for structured outputs; LLM-as-a-judge by means of Amazon Bedrock Mannequin Analysis for subjective high quality evaluation throughout helpfulness, correctness, and security; and human analysis for high-stakes use instances the place automated metrics may miss nuances. Combine these evaluations into your steady integration and steady supply (CI/CD) pipeline to catch regressions early.
For manufacturing deployment, implement a phased rollout. Begin with shadow testing to check Nova 2 outputs towards Nova 1 with out affecting customers. Progress to A/B testing with a small proportion of visitors to measure enterprise impression. Use canary releases or blue/inexperienced deployments to take care of rollback functionality if points come up.
Constructing responsibly with Nova 2
Amazon Nova 2 fashions are developed with security, safety, and belief as core priorities all through the mannequin growth lifecycle, with built-in content material moderation aligned to the AWS Acceptable Use Coverage. AWS additionally gives uncapped mental property indemnity protection for outputs of usually out there Amazon Nova fashions. Whereas AWS offers these foundational safeguards, you share the duty for accountable deployment. Throughout your migration, consider Nova 2 outputs for accuracy and appropriateness in your particular use instances, notably for functions that floor outcomes straight to finish customers or inform consequential selections. For prime-stakes workloads, implement acceptable human oversight and use-case-specific guardrails similar to Amazon Bedrock Guardrails. For detailed steerage on Amazon Nova’s accountable AI method and proposals, see Accountable use of Amazon Nova.
Migration guidelines
Use this guidelines to plan and execute your Nova 1 to Nova 2 migration.
- Determine your present Nova 1 mannequin and goal Nova 2 mannequin (Lite → 2 Lite, Professional → 2 Lite)
- Request mannequin entry within the Amazon Bedrock console for every Area that you simply want
- Replace mannequin IDs in your code (for instance, us.amazon.nova-lite-v1:0 → us.amazon.nova-2-lite-v1:0)
- Configure shopper timeout to three,600 seconds for prolonged considering workloads
- If utilizing excessive effort reasoning, take away temperature, topP, and topK parameters
- Replace IAM insurance policies if utilizing built-in instruments (InvokeTool permission required)
- Run analysis exams evaluating Nova 1 and Nova 2 outputs in your manufacturing prompts
- Deploy utilizing shadow testing or canary launch earlier than full rollout
Cleansing up
When you created take a look at assets whereas following this information, delete them to keep away from ongoing prices. This consists of:
- Any take a look at Amazon Bedrock invocations (these are charged per token, so no cleanup wanted except you provisioned throughput)
- Check IAM insurance policies or roles created particularly for Nova 2 migration testing
- Any Amazon CloudWatch Logs generated throughout testing, when you now not want them for debugging
When you provisioned devoted throughput for testing the Nova 2 mannequin, delete the provisioned throughput within the Amazon Bedrock console to cease incurring prices.
Conclusion
Migrating from Nova 1 to Nova 2 on Amazon Bedrock is easy for many workloads. Replace your mannequin IDs, regulate parameters when you plan to make use of excessive effort reasoning, and run analysis exams in your manufacturing prompts earlier than full deployment. The core API construction stays the identical, so present Converse API integrations require minimal code modifications.
Nova 2 brings capabilities that may enhance your functions with out architectural modifications: prolonged considering for complicated reasoning, built-in instruments like net grounding and code interpreter, and an expanded context window for longer paperwork. To get began, request entry to the Nova 2 mannequin within the Amazon Bedrock console and run a comparability take a look at towards your present Nova 1 deployment. For detailed documentation, see the Amazon Nova Consumer Information.
Acknowledgement
Particular due to Sneha Venkateswaran, Sharon Li and Jean Farmer for his or her contribution.
Concerning the authors
Adewale Akinfaderin
Adewale is a Sr. Knowledge Scientist–Generative AI, Amazon Bedrock, the place he contributes to innovative improvements in foundational fashions and generative AI functions at AWS. His experience is in reproducible and end-to-end AI/ML strategies, sensible implementations, and serving to international prospects formulate and develop scalable options to interdisciplinary issues. He has two graduate levels in physics and a doctorate in engineering.
Veda Raman
Veda is a Sr Options Architect for Generative AI for Amazon Nova and Agentic AI at AWS. She helps prospects design and construct Agentic AI options utilizing Amazon Nova fashions and Bedrock AgentCore. She beforehand labored with prospects constructing ML options utilizing Amazon SageMaker and in addition as a serverless options architect at AWS.
Wrick Talukdar
Wrick is a expertise chief and Senior Generative AI Specialist at AWS, targeted on multimodal basis fashions and agentic AI. He’s the bestselling creator of Constructing Agentic AI Methods and Generative AI Ethics, Privateness, and Safety, serves as Chair of the IEEE Expertise & Intelligence group for 2026–2027, and regularly speaks at main international boards together with AWS re:Invent, ICCE, CERAWeek, and ADIPEC. Outdoors work, he enjoys writing and birding pictures.
