OpenAI open weight fashions now out there on AWS

AWS is dedicated to bringing you essentially the most superior foundation fashions (FMs) within the trade, constantly increasing our choice to incorporate groundbreaking fashions from main AI innovators so that you just at all times have entry to the newest developments to drive your enterprise ahead.

Today, I’m glad to announce the supply of two new OpenAI fashions with open weights in Amazon Bedrock and Amazon SageMaker JumpStart. OpenAI gpt-oss-120b and gpt-oss-20b fashions are designed for textual content era and reasoning duties, providing builders and organizations new choices to construct AI purposes with full management over their infrastructure and knowledge.

These open weight fashions excel at coding, scientific evaluation, and mathematical reasoning, with efficiency similar to main options. Both fashions help a 128K context window and supply adjustable reasoning ranges (low/medium/excessive) to match your particular use case necessities. The fashions help exterior instruments to boost their capabilities and can be utilized in an agentic workflow, for instance, utilizing a framework like Strands Agents.

With Amazon Bedrock and Amazon SageMaker JumpStart, AWS provides you the liberty to innovate with entry to a whole lot of FMs from main AI corporations, together with OpenAI open weight fashions. With our complete number of fashions, you possibly can match your AI workloads to the right mannequin each time.

Through Amazon Bedrock, you possibly can seamlessly experiment with completely different fashions, combine and match capabilities, and swap between suppliers with out rewriting code—turning mannequin selection right into a strategic benefit that helps you constantly evolve your AI technique as new improvements emerge. These new fashions can be found in Bedrock through an OpenAI-compatible endpoint. You can level the OpenAI SDK to this endpoint or use the Bedrock InvokeModel and Converse API.

With SageMaker JumpStart, you possibly can rapidly consider, evaluate, and customise fashions to your use case. You can then deploy the unique or the personalized mannequin in manufacturing with the SageMaker AI console or utilizing the SageMaker Python SDK.

Let’s see how these work in apply.

Getting began with OpenAI open weight fashions in Amazon Bedrock
In the Amazon Bedrock console, I select Model entry from the Configure and study part of the navigation pane. Then, I navigate to the 2 listed OpenAI fashions on this web page and request entry.

Now that I’ve entry, I take advantage of the Chat/Test playground to check and consider the fashions. I choose OpenAI because the class after which the gpt-oss-120b mannequin.

Using this mannequin, I run the next pattern immediate:

A household has $5,000 to save lots of for his or her trip subsequent 12 months. They can place the cash in a financial savings account incomes 2% curiosity yearly or in a certificates of deposit incomes 4% curiosity yearly however with no entry to the funds till the holiday. If they want $1,000 for emergency bills through the 12 months, how ought to they divide their cash between the 2 choices to maximise their trip fund?

This immediate generates an output that features the chain of thought used to supply the outcome.

I can use these fashions with the OpenAI SDK by configuring the API endpoint (base URL) and utilizing an Amazon Bedrock API key for authentication. For instance, I set this atmosphere variables to make use of the US West (Oregon) AWS Region endpoint (us-west-2) and my Amazon Bedrock API key:

export OPENAI_API_KEY=""
export OPENAI_BASE_URL="https://bedrock-runtime.us-west-2.amazonaws.com/openai/v1"

Now I invoke the mannequin utilizing the OpenAI Python SDK.

from openai import OpenAI

shopper = OpenAI()

response = shopper.chat.completions.create( 
    messages=[{ "role": "user", "content": "Tell me the square root of 42 ^ 3" }],
    mannequin="openai.gpt-oss-120b-1:0",
    stream=False
)

for merchandise in response:
    print(merchandise)

I save the code (test-openai.py file), set up the dependencies, and run the agent regionally:

pip set up openai
python test-openai.py

To construct an AI agent, I can select any framework that helps the Amazon Bedrock API or the OpenAI API. For instance, right here’s the beginning code for Strands Agents utilizing the Amazon Bedrock API:

from strands import Agent
from strands.fashions import BedrockModel

bedrock_model = BedrockModel(
    model_id="openai.gpt-oss-120b-1:0",
    region_name="us-west-2",
    streaming=False
)

agent = Agent(
    mannequin=bedrock_model
)

agent("Tell me the sq. root of 42 ^ 3")

I save the code (test-strands.py file), set up the dependencies, and run the agent regionally:

pip set up strands-agents
python test-strands.py

When I’m glad with the agent, I can deploy in manufacturing utilizing the capabilities supplied by Amazon Bedrock AgentCore, together with a completely managed serverless runtime and reminiscence and identification administration.

Getting began with OpenAI open weight fashions in Amazon SageMaker JumpStart
In the Amazon SageMaker AI console, you should use OpenAI open weight fashions within the SageMaker Studio. The first time I do that, I must arrange a SageMaker area. There are choices to set it up for a single consumer (less complicated) or a company. For these exams, I take advantage of a single consumer setup.

In the SageMaker JumpStart mannequin view, I’ve entry to an in depth description of the gpt-oss-120b or gpt-oss-20b mannequin.

I select the gpt-oss-20b mannequin after which deploy the mannequin. In the subsequent steps, I choose the occasion sort and the preliminary occasion depend. After a couple of minutes, the deployment creates an endpoint that I can then invoke in SageMaker Studio and utilizing any AWS SDKs.

To study extra, go to GPT OSS fashions from OpenAI are actually out there on SageMaker JumpStart within the AWS Artificial Intelligence Blog.

Things to know
The new OpenAI open weight fashions are actually out there in Amazon Bedrock within the US West (Oregon) AWS Region, whereas Amazon SageMaker JumpStart helps these fashions in US East (Ohio, N. Virginia) and Asia Pacific (Mumbai, Tokyo).

Each mannequin comes geared up with full chain-of-thought output capabilities, offering you with detailed visibility into the mannequin’s reasoning course of. This transparency is especially beneficial for purposes requiring excessive ranges of interpretability and validation. These fashions provide the freedom to switch, adapt, and customise them to your particular wants. This flexibility lets you fine-tune the fashions to your distinctive use circumstances, combine them into your present workflows, and even construct upon them to create new, specialised fashions tailor-made to your trade or software.

Security and security are constructed into the core of those fashions, with complete analysis processes and security measures in place. The fashions preserve compatibility with the usual GPT-4 tokenizer.

Both fashions can be utilized in your most popular atmosphere, whether or not that’s by means of the serverless expertise of Amazon Bedrock or the intensive machine studying (ML) improvement capabilities of SageMaker JumpStart. For details about the prices related to utilizing these fashions and providers, go to the Amazon Bedrock pricing and Amazon SageMaker AI pricing pages.

To study extra, see the parameters for the fashions and the chat completions API within the Amazon Bedrock documentation.

Get began immediately with OpenAI open weight fashions on AWS within the Amazon Bedrock console or in Amazon SageMaker AI console.

– Danilo