DeepSeek‑V3.1 Launches on Amazon Bedrock: Fully Managed Model with Dual Reasoning Modes

In March, Amazon Web Services became the first provider to offer the DeepSeek‑R1 model as a serverless service on Amazon Bedrock. Today DeepSeek‑V3.1 is released as a fully managed foundation model on Bedrock, supporting two inference modes: a “thinking” mode that outputs step‑by‑step reasoning and a “non‑thinking” mode that returns answers directly.

According to DeepSeek official data, compared with the DeepSeek‑R1‑0528 version, DeepSeek‑V3.1’s thinking mode delivers answer quality that is comparable or better, stronger multi‑step reasoning on complex search tasks, and a significant boost in reasoning efficiency.

After training optimisation, the model shows markedly better performance on tool‑calling and agent tasks, supports over 100 languages at near‑native proficiency—especially improving low‑resource language handling—and reduces hallucinations while providing clearer decision traceability.

Core application scenarios:

Code generation: excels in programming benchmarks and code‑agent tasks, making it suitable for automated code generation, debugging, and software‑engineering workflows.

Agentic AI tools: enhanced tool‑calling ability, structured tool usage, and support for code and search agents, enabling robust autonomous AI systems.

Enterprise‑grade applications: integrated into various chat and productivity tools, offering multilingual interaction and cultural adaptability for global deployments.

When deploying a publicly available model, users should consider data‑privacy requirements, monitor for bias, and leverage Amazon Bedrock’s security features and Guardrails, as well as the Bedrock model evaluation tools for comparative testing.

To start using DeepSeek‑V3.1, request access in the Bedrock console, then invoke the model via the console, the AWS CLI, or the Amazon SDK. The model provides both InvokeModel and Converse APIs with extensive code examples covering multiple programming languages.

The model is currently available in the following regions: US West (Oregon), Asia Pacific (Tokyo), Asia Pacific (Mumbai), Europe (London), and Europe (Stockholm). Future region updates are listed in the full region list.

Example prompt for architecture design:

Outline the high‑level architecture for a scalable URL shortener service like bit.ly. Discuss key components such as API design, database choice (SQL vs. NoSQL), how the redirect mechanism works, and how you would generate unique short codes.

For detailed inference parameters and response specifications, refer to the DeepSeek model documentation on the Amazon Bedrock site.