What AWS announced
AWS What’s New reports that DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash are available through Amazon SageMaker JumpStart. This is an AWS vendor announcement about access to DeepSeek models through an AWS deployment service—not a research paper, community post, or independent performance assessment. DeepSeek is the model provider; AWS is the publisher of this availability announcement.
According to the supplied announcement, all three are open-weight Mixture-of-Experts (MoE) language models released under the MIT license. AWS says customers can deploy them through the SageMaker JumpStart model catalog in the SageMaker console or through the SageMaker Python SDK.
The supplied text labels the announcement November 11, 2026, while the source URL contains a January 2026 path. Those fields are inconsistent, so the publication date should be checked against the original page before being used as a release-history reference.
How AWS describes the models
The announcement positions the models for different workloads. The specifications below are AWS’s reported figures, not independently verified measurements:
- DeepSeek-V4-Pro: AWS describes a 1.6-trillion-parameter MoE model that activates 49 billion parameters per token and supports a one-million-token context window. Its stated focus is complex reasoning, long-document analysis, and demanding text generation.
- DeepSeek-V4-Flash: AWS reports 284 billion total parameters, with 13 billion activated per token and the same one-million-token context window. The announcement positions it for latency-sensitive and cost-conscious language workloads.
- DeepSeek-V4.1-Flash: AWS describes native multimodal support and a Causal Encoder-Decoder architecture with a 552-billion-parameter backbone. It reports eight billion active parameters during prefill, 16 billion during decoding, a one-million-token context window, and a maximum output of 384,000 tokens.
For V4.1-Flash, AWS also attributes a smaller global key-value cache footprint to Compressed Sparse Attention 2 and FP4 KV caching. Its comparison is roughly one-quarter of V4-Flash’s footprint. That is a specific architectural claim, not evidence of an equivalent reduction in a customer’s AWS bill.
Business impact: more candidates, not proven savings
SaveAWS analysis: The practical opportunity is a broader set of models to evaluate within a SageMaker-based deployment workflow. Teams working on document analysis, coding assistants, or agent workflows could compare these options against their existing models using the same business tasks.
However, active parameter counts do not determine total deployment cost on their own. Model-weight memory, serving configuration, request concurrency, input lengths, output lengths, and utilization can all affect infrastructure requirements and economics. A large supported context window also does not establish acceptable latency or answer quality at that length.
The announcement’s descriptions of frontier-level performance and improved efficiency are vendor claims. The supplied material includes no benchmark methodology, workload-specific cost results, or independent validation. It therefore does not support a savings percentage or a conclusion that one model will be cheaper for every deployment.
What to verify before deployment
The supplied announcement does not list supported AWS Regions, instance requirements, prices, or deployment quotas. Avoid assuming that catalog availability means every account can deploy every model in every Region.
Before committing production traffic:
- Confirm the exact model entry, supported Region, deployment configuration, and account prerequisites in SageMaker JumpStart.
- Review the model license, model documentation, and applicable service terms with the relevant internal owners.
- Run representative evaluations for answer quality, coding reliability, multimodal behavior where applicable, and failure handling.
- Measure latency, throughput, memory requirements, and total deployment cost at realistic concurrency and context lengths.
- Set spending controls and remove test deployments when evaluation is complete.
For agent workflows, assess end-to-end task completion rather than token generation alone. Retries, tool calls, and unsuccessful runs can change both operating cost and business value.
Original source
Publisher: AWS What’s New. Read the original AWS announcement. This article distinguishes AWS’s reported specifications from SaveAWS’s evaluation guidance; it does not independently certify the models’ performance or availability.
AWS What’s New
DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash models now available on Amazon SageMaker JumpStart · 2026-11-11