Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- This is an availability update for three models in AWS SageMaker JumpStart. It does not mean every DeepSeek product or cloud platform changed at the same time.
- AWS positions V4-Pro for complex reasoning, V4-Flash for lower compute cost and latency-sensitive work, and V4.1-Flash for agentic, coding and vision tasks. These are AWS descriptions, not PuppyIP benchmark rankings.
- All three list a 1M-token context window. V4.1-Flash also lists up to 384K output tokens. Context, maximum output and per-request limits are separate constraints.
- The release note gives no invocation IDs, Region eligibility, instance specifications or numeric prices. Check the JumpStart catalog and current billing information for your own account before deployment.
Start with what AWS made available
In its November 11, 2026 release note, AWS announced that DeepSeek-V4-Pro, DeepSeek-V4-Flash and DeepSeek-V4.1-Flash are available through Amazon SageMaker JumpStart. This guide covers the AWS catalog and deployment path. It is not an announcement that DeepSeek’s direct API, other cloud platforms or the entire DeepSeek product line changed at the same time.
AWS says customers can use the SageMaker JumpStart model catalog or the SageMaker Python SDK to deploy the models to an AWS account. That confirms AWS has announced platform availability; it does not establish that every AWS account, organization policy or Region can deploy each model immediately. Check the model card visible to the account you plan to use.
Compare the three models by workload
DeepSeek-V4-Pro is the flagship model for complex reasoning and high-quality text generation. It uses a mixture-of-experts (MoE) architecture, which activates only part of the model for each token, a basic unit of content processed by a model. AWS lists 1.6T total parameters, with 49B active per token and a 1M-token context window. It may be a candidate for multi-step reasoning or long documents, but parameter scale alone does not guarantee answer quality.
DeepSeek-V4-Flash has 284B parameters, activates 13B per token and offers the same 1M-token context. AWS positions it as a lower-compute-cost option for latency-sensitive and cost-conscious workloads.
DeepSeek-V4.1-Flash uses a 552B Causal Encoder-Decoder backbone, activating 8B parameters while processing input (prefill) and 16B while generating output (decode). AWS highlights it for agentic, coding and vision tasks and describes it as natively multimodal. These capabilities and positioning come from the release note, not a side-by-side test on one dataset.
Read context, output and cache figures separately
All three models list a 1M-token context window. AWS also lists up to 384K output tokens for V4.1-Flash. That output maximum is a separate specification for that model; it should not be added to the 1M context figure or treated as a promise that every request can use both limits in full. Plan input, output and truncation against the model card and actual interface limits.
V4.1-Flash uses Compressed Sparse Attention 2 (CSA2) and FP4 KV caching, which retains attention information for subsequent generation. AWS says its global KV-cache footprint is about one quarter of V4-Flash’s and cites this as a cost-efficiency benefit for agentic, coding and vision workloads. This describes cache footprint; it is not a dollar quote, an overall throughput test or a guarantee that every task will cost less.
Check the actual deployment details in the catalog
The AWS announcement gives reader-facing model names, but no exact SageMaker invocation IDs, endpoint parameters or shared request format. Open the JumpStart model card or the documentation for the relevant SDK, then record the current model identifier and invocation method. Do not copy a product name from this guide as though it were an API parameter.
The announcement also does not list supported Regions, account eligibility, instance types or numeric rates. For the account and Region you intend to use, confirm whether each model is selectable, what compute resources it requires and how it is billed. The MIT license applies to the model weights; it does not establish that hosted inference is free. The release note itself contains no price list.
Evaluate with a small, representative workload
Start with representative prompts and input samples. Run the same tasks on your candidate models, then compare output quality, response time, failures and the billing information shown for your actual configuration. Record the model ID, version, Region, input length and output limit used so results from different settings are not mistaken for model differences.
This is a practical comparison method, not an AWS deployment requirement or a benchmark result claimed by this guide. If a workload depends on vision input, long context or tool use, first verify the capabilities and SDK behavior documented for the selected model, then decide whether to expand traffic.
Sources
Frequently Asked Questions
Do the MoE total or active parameter counts tell me which SageMaker instance to choose?
No. The release note gives model parameter counts, but no matching instance types, memory requirements or deployment topology. Use the current JumpStart model card for deployment options, then verify resources and cost with your actual account configuration.
Does “native multimodal” mean V4.1-Flash accepts audio and video input?
The announcement alone does not establish that. AWS highlights vision tasks and calls the model natively multimodal, but the release note does not list every input modality or format. Check the current model card and interface documentation for this model.