With general release of Amazon Nova Premier, most powerful model for intricate jobs and a model distillation instructor, they growing the Amazon Nova family of foundation models that were unveiled at AWS re:Invent.
The Amazon Nova understanding models that are already offered in Amazon Bedrock are joined by Nova Premier. Premier can process input text, photos, and videos (but not audio), just like Nova Lite and Pro. With its sophisticated features, Nova Premier is excellent at difficult jobs requiring multi-step planning, accurate execution across a variety of tools and data sources, and a thorough grasp of context. Nova Premier can handle very long texts or big code bases since it has a context length of one million tokens.
For your particular requirements, you may construct very powerful, reasonably priced, and low-latency versions of Nova Pro, Lite, and Micro using Nova Premier with Amazon Bedrock Model Distillation. For instance, it distilled Nova Pro for intricate tool selection and API calls using Nova Premier. With the speed and cost advantages of Nova Pro, the distilled Nova Pro regularly equalled the teacher’s performance and had a 20% greater accuracy for API invocations than the basic model.
Amazon Nova Premier benchmark evaluation
Amazon assessed Nova Premier using a variety of benchmarks related to agentic workflows, visual intelligence, and text intelligence. According to the chart below, which compares the Nova Premier against 17 benchmarks, it is the most competent model in the Nova line.
Comparing Nova Premier to other models in the same intelligence tier, it is equal to or better on around half of these criteria, and it is similar to the finest non-reasoning models in the industry. The technical report contains specifics on these assessments.
For its intelligence category, Nova Premier is also Amazon Bedrock‘s quickest and most economical model.
Additionally, Amazon may be used as a teaching model for distillation, meaning that its sophisticated features for a particular use case can be transferred to smaller, quicker, and more effective models such as Nova Pro, Micro, and Lite for production deployments.
Using Amazon Nova Premier
You must first ask for access to the model in the Amazon Bedrock dashboard in order to begin using Nova Premier. Locate Nova Premier in the navigation window, choose Model access, and toggle access.
Once you have access, you can use Nova Premier by entering a collection of messages between the user and the assistant into the Amazon Bedrock Converse API. Videos, pictures, and text can all be included in messages. This is an illustration of a simple call with the AWS SDK for Python (Boto3):
This illustration demonstrates how Nova Premier may offer thorough justifications for intricate technical queries. However, Premier’s true strength lies in its capacity to manage complex processes.
Multi-agent collaboration use case
In order to demonstrate how Nova Premier operates a multi-agent cooperation architecture for investment research, let’s examine a more intricate situation.
Finding pertinent data sources for particular investments, extracting the necessary information from those sources, and combining the data into actionable insights are the usual steps in the equity research process. When working with several financial products, such as stock indexes, individual stocks, and currencies, this procedure gets more complicated.
With Nova Premier powering the supervisor agent that coordinates the entire process, it can create this kind of application on Amazon Bedrock through multi-agent collaboration. For instance, “What are the emerging trends in renewable energy investments?” is an example of an initial question that the supervisor agent analyses, breaks down into logical stages, chooses which specialised subagents to engage, and synthesizes the final response.
To developed a system for this situation that consists of the following elements:
- A Nova Premier-powered supervisor agent
- Nova Pro powers a number of specialised subagents that concentrate on various financial data sources.
- Tools that link to market analysis programs, financial databases, and other pertinent information sources
The supervisor agent enabled by Nova Premier performs the following actions when it ask a question concerning new developments in investments in renewable energy:
- Determines the underlying subjects and sources to cover by analysing the query.
- Chooses the right subagents for those subjects and sources.
- Technical analysis, market mood data, and pertinent economic information are retrieved by each subagent.
- The supervisor agent compiles this data into an extensive report that a financial professional can review.
By using Nova Premier in a multi-agent collaboration architecture, the financial professional’s work is streamlined and their investment analysis is developed more quickly. This situation is illustrated visually in the video that follows.
The main benefit of employing Nova Premier for the supervisor position is its precision in managing intricate processes, ensuring that the appropriate data sources are consulted in the best possible order and that each subagent receives the appropriate input for their task, leading to insights of superior quality.
Multi-agent collaboration with model distillation
Even though Nova Premier is the most accurate model in its family, you may wish to minimise latency and expenses in production settings. This is where Nova Premier’s power as a distillation teaching approach gets intriguing. For this particular investment research use case, it can modify Nova Micro from the outcomes of Nova Premier using Amazon Bedrock Model Distillation.
By having a teacher model provide the necessary outputs, model distillation streamlines the data collecting process and allows you to generate high-quality training data, in contrast to traditional fine-tuning, which necessitates human input and labelled examples.
The following steps are involved in distilling a model:
- Using input and output from Nova Premier runs over a variety of financial instruments to create synthetic training data
- Train a customised Nova Micro utilising special fine-tuning instruments using this data.
- Assessing the customised Micro model’s performance and latency differences
- Using customised Micro model as production supervisor agent
You may leverage invocation logs for data preparation and further simplify the process with Amazon Bedrock. In order to accomplish this, you must enable model invocation logging and configure an Amazon Simple Storage Service (Amazon S3) bucket as the log destination.