AWS adds fully managed AI models: Qwen3 and DeepSeek-V3.1

Amazon Web Services Inc. today announced the addition of fully managed open-weight models Qwen3 and DeepSeek-V3.1 to its AI model portfolio.

The new models offer greater flexibility to customers that rely on the Amazon Bedrock generative AI service to meet their evolving business needs.

Open-weight models provide increased transparency for developers regarding model weights, which makes it easier to customize models for specific use cases. The new open-weight models in Amazon Bedrock join others from leading developers, including Meta Platforms Inc., Mistral AI and OpenAI.

There is a veritable cornucopia of models that have strengths in different areas. Qwen3, from Alibaba, delivers model options for sophisticated coding and general reasoning and the DeepSeek-V3.1 model delivers exceptional performance across math, coding and agentic tasks. Qwen3 is the first fully managed Qwen model family in the Amazon Bedrock portfolio.

While the models are freely available, by using these models in Bedrock, customers can take advantage of Amazon’s enterprise-grade security, including data encryption and strict access controls, which help maintain data privacy and regulatory compliance. Customers retain full control over their data, as AWS does not share model input and output data with model providers, and it is not used to improve the base models.

Delivering customer choice in more global regions

Prior to the announcement, I spoke with Shaown Nandi, director of technology for AWS, about the value the new models will bring to customers. Nandi, who was chief information officer of News Corp.’s Dow Jones unit before joining AWS six years ago, said AWS will launch the new models in key global markets such as Asia, Latin America, Europe and North America.

He said large general-purpose AI models can be too big for the narrow use cases common across many enterprises. “You want a smaller model, or a cheaper model, and that’s OK because of the variety of use cases,” Nandi said. “What we’re seeing with open-weight models, specifically, is a cost advantage, a choice advantage. But also, with models like Llama, on which AWS supports model distillation, you’re seeing customers in Bedrock able to train this model down to a smaller size and retain much of the accuracy at a cost that could be as much as 30 times cheaper to run after it’s been distilled.”

“Whether it’s selecting a narrower model, distilling the model, or simply avoiding the higher licensing costs of some proprietary models — like agentic use cases — that’s where these open-weight models start to really shine,” he said.

Nandi said customers in Latin America and parts of Asia have a particular interest in being able to tune models for their local needs, which can be easier to achieve with open-weight models. “I see a lot of demand in international markets and from startups and other companies in the U.S., as well,” he said.

Open-weight models also deliver the speed and flexibility organizations demand. “What’s unique with open-weight models is the ability to fine-tune and customize them,” Nandi said. “We’re seeing customers experimenting with these models, whether it’s distilling them or fine-tuning them in different sizes and effectively building something that looks like their own SLM for their industry or business.”

Qwen3 capabilities

AWS said its customers now have access to four new open-weight models from the Qwen3 family. These multilingual models can plan multistep workflows, integrate with tools and APIs, and handle long context windows within a task. The two general-purpose models provide both “thinking” and “non-thinking” reasoning modes.

In addition, the announcement said, if the Qwen3 models “were people,” they could “speak dozens of languages fluently and share an encyclopedic knowledge of diverse subjects, from explaining scientific concepts to writing creative stories.”

DeepSeek-V3.1 offerings

The strength of the DeepSeek-V3.1 model is hybrid reasoning capabilities that balance fast responses with deep, transparent thinking by enabling customers to toggle between modes depending on the type of problems they are trying to solve.

Also, the models are very energy-efficient, since they rarely turn basic queries into long, drawn out discussions while maintaining a high level of expertise for making strategic decisions. Also, the models explain their thinking clearly, making it easier to understand how it reached its recommendations.

Leveraging customer feedback

There’s no crystal ball that tells AWS — or any model developer — what models to launch in which markets. So, they listen to customers, look at usage, and make educated guesses about deployment and update plans.

“That last piece is important,” said Nandi. “We want to fill in gaps. We want customers to have the full weight of choice. And there are so many new agentic use cases coming up right now. We’re constantly under pressure to add more models.” Today, AWS offers hundreds of models and is expanding the regions in which it adds new models.

Another source of customer feedback is Bedrock’s Model Evaluation tool. “It uses a large language model as a judge,” Nandi said. “Based on the parameters you input, it tells you which model makes the most sense for you. That’s our scalable method of giving customers automated feedback in Bedrock. It’s been a game changer for customers.”

These new models, and the strategy of bringing Amazon Bedrock models to customers in more global regions, make good business sense for AWS. They should provide its current and future customers with ever more capable foundation model solutions to grow their businesses.

Zeus Kerravala is a principal analyst at ZK Research, a division of Kerravala Consulting. He wrote this article for SiliconANGLE.

Photo: Robert Hof/SiliconANGLE