Back to AWS content
AWS What's New

DeepSeek-V4 Models Now on SageMaker JumpStart

AWS adds DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash to SageMaker JumpStart, offering new open-weight MoE models for reasoning, coding, and agentic tasks.

5 min read·Curated & commentary by AWS News Bot
sagemakerllmaijumpstartdeepseek

DeepSeek's DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers.

In practice: This launch adds three new open-weight MoE models to SageMaker JumpStart, each tuned for different efficiency and performance trade-offs.

The news, in plain English

“These models expand SageMaker JumpStart’s foundation model portfolio with new open-weight MoE options for reasoning, coding, and agentic workloads.”

AWS has added DeepSeek’s DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash models to Amazon SageMaker JumpStart. All three are Mixture-of-Experts (MoE) architectures released under the MIT license. They target high-performance reasoning, coding, and agentic use cases while maintaining efficiency through partial parameter activation per token.

Technical context

The models share several key traits that distinguish them in the foundation model landscape:

  • MoE architecture: Each model activates only a subset of its total parameters per token. This design reduces both compute and memory requirements during inference compared to dense models of similar total size.
  • Open weights: All three models are released under the MIT license, allowing full access to model weights and architectures for customization, research, and commercial use.
  • Large context windows: Every model supports a 1M-token context window. This scale enables complex prompt engineering, long-document analysis, and multi-step reasoning tasks that would otherwise require careful chunking or summarization.

Specific model details highlight their architectural differences:

  • DeepSeek-V4-Pro: A 1.6T-parameter MoE model activating 49B parameters per token during inference. This model emphasizes maximum performance for demanding workloads.
  • DeepSeek-V4-Flash: A 284B-parameter MoE model activating 13B parameters per token. It balances performance and efficiency for scale-intensive applications.
  • DeepSeek-V4.1-Flash: A 552B-parameter backbone that switches between 8B parameters during prefill (when the model processes input tokens) and 16B parameters during decode (when it generates output tokens). It also supports a maximum output token length of 384K, extending its utility for generation-heavy tasks.

Where this fits in practice

These models are likely to be useful in scenarios that require specialized capabilities beyond general-purpose language understanding:

  • Complex reasoning: Tasks that involve multi-step logic, mathematical computation, hypothesis testing, or chain-of-thought reasoning where traditional models struggle with accuracy or resource consumption.
  • Coding assistance: Support for code generation, debugging, refactoring, or comprehension across multiple programming languages, potentially improving developer productivity and code quality.
  • Agentic workflows: Building systems where AI agents plan, execute, and reflect on sequences of actions—such as automated troubleshooting, workflow orchestration, or interactive decision support.

Watch out: While the models share similar design goals, their performance characteristics and efficiency trade-offs differ significantly. Selecting the right model will depend on the specific requirements of each workload, such as the need for maximum performance versus minimal inference cost or latency. For example, DeepSeek-V4-Pro may be ideal for accuracy-critical tasks, while DeepSeek-V4-Flash could better serve high-throughput applications.

Adoption path

AWS customers can access these models through SageMaker JumpStart, which streamlines discovery and deployment. The models can be deployed using the SageMaker SDK or AWS CLI, though the fact pack does not include specific implementation steps, code examples, or configuration details. Users will need to refer to the SageMaker documentation or DeepSeek’s official resources for deployment guidance. This means teams must invest time in learning the particular integration patterns and infrastructure requirements for each model variant.

Cost and operations

The fact pack does not provide any details on pricing, quotas, or regional availability. As a result, it is not possible to determine the cost implications or any potential constraints on model usage. Users should expect to incur standard SageMaker inference and training costs, but exact figures—including per-token or per-invocation pricing—are unavailable at this time. This absence of financial clarity makes budgeting for production workloads challenging.

Security and reliability

Security and reliability details are not provided in the source material. Users should apply standard AWS security practices for SageMaker, such as IAM role management, VPC isolation, and encryption at rest and in transit. No specific vulnerabilities or reliability guarantees are mentioned in the announcement. Teams deploying these models must therefore rely on general SageMaker security configurations and monitor performance through their own observability tools.

Limits and trade-offs

Several key limitations are evident based on the information provided:

  • Missing pricing and quota details: Without this information, planning for production use is challenging. Teams cannot accurately forecast operational expenses or anticipate request limits that might affect scalability.
  • No benchmark data: Performance comparisons to other models or hardware are not provided, making it difficult to assess relative strengths or weaknesses in real-world scenarios.
  • Regional availability unknown: It is unclear which AWS regions host these models, potentially impacting latency, data residency requirements, and access for certain customers.

These gaps make it difficult to fully assess the operational impact or cost-effectiveness of deploying these models at scale. Teams should treat initial adoption as experimental until more comprehensive data becomes available.

Who should act

  • AI researchers and developers: Those exploring open-weight MoE models for reasoning, coding, or agentic tasks may want to experiment with these models to evaluate their suitability for specific use cases.
  • Teams with SageMaker expertise: Users already familiar with SageMaker workflows can quickly add these models to their pipelines, pending clarity on pricing and regional support. Their existing infrastructure and knowledge reduce time-to-value for prototyping.

In practice: Teams should start small, using these models for specific tasks where their unique capabilities—such as long context windows or efficient parameter usage—provide clear advantages. Early experiments can inform larger investments once pricing and performance data become available.

Bottom line

The addition of DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash to SageMaker JumpStart expands AWS’s portfolio of foundation models with new open-weight MoE options. While the models show promise for complex reasoning, coding, and agentic workloads, the lack of pricing, quota, and benchmark information means that adoption should proceed with caution. Users should monitor AWS for additional details before committing to production use. Until clearer operational metrics emerge, these models are best suited for research, proof-of-concept, and low-risk pilot projects.

Sources

DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash models now available on Amazon SageMaker JumpStart