DeepSeek V4 on AWS: SageMaker JumpStart, Bedrock status, price and how to deploy

DeepSeek V4 Pro, V4 Flash and V4.1 Flash are on SageMaker JumpStart, not Bedrock. Specs, API price comparison and how to deploy on AWS.

Sanchana Charles16 November 202610 min read0 viewsAWS
DeepSeek V4 on AWS: SageMaker JumpStart, Bedrock status, price and how to deploy

DeepSeek V4 is now on AWS, but not where most people expect. On 11 November 2026 AWS added DeepSeek-V4-Pro, DeepSeek-V4-Flash and DeepSeek-V4.1-Flash to Amazon SageMaker JumpStart, which means you deploy the open weights onto GPU instances in your own account and pay for those instances by the hour. They are not on Amazon Bedrock's serverless, pay-per-token model list: as of 16 November, Bedrock's newest DeepSeek model is still DeepSeek V3.2.

Below: what each V4 model is, how JumpStart pricing compares with DeepSeek's own API, how to deploy one, and a decision table for picking the right route.

What did AWS announce for DeepSeek V4?

AWS's What's New post, dated 11 November 2026, says three DeepSeek models "are now available on Amazon SageMaker JumpStart". AWS describes them as "open-weight Mixture-of-Experts (MoE) language models, released under the MIT license". The key specs, as AWS lists them:

Model Total parameters Active per token Context window AWS says it's for
DeepSeek-V4-Pro 1.6T 49B 1M tokens Complex reasoning, long documents, "maximum model capacity"
DeepSeek-V4-Flash 284B 13B 1M tokens "Latency-sensitive and cost-conscious production deployments"
DeepSeek-V4.1-Flash 552B backbone 8B (prefill) / 16B (decode) 1M tokens, 384K max output Agentic work, coding and vision (native multimodal)

AWS adds that V4.1-Flash uses a "Causal Encoder-Decoder (CED) architecture" plus "Compressed Sparse Attention 2 (CSA2) and FP4 KV caching", which it says cut the KV cache footprint to "roughly 1/4 of V4-Flash". Those are AWS's and DeepSeek's claims; we haven't benchmarked them.

DeepSeek's own Hugging Face model card for V4-Pro confirms the 1.6T total / 49B active split, the one-million-token context and the MIT licence for the weights, and says the weights use mixed "FP4 + FP8" precision.

Despite the "2026/01" in the URL of AWS's announcement, the page itself says "Posted on: Nov 11, 2026".

Is DeepSeek V4 available on Amazon Bedrock?

No, not as a serverless Bedrock model, at the time of writing. Bedrock's official "Models at a glance" page lists three DeepSeek models: DeepSeek V3.2, DeepSeek-V3.1 and DeepSeek-R1. No V4 model appears there.

That's the single most common point of confusion. Google's autocomplete is full of searches like "deepseek v4 bedrock pricing" and "deepseek v4 pro aws bedrock", but those searches have no official answer yet. Some third-party pages have claimed V4 was in a Bedrock preview earlier in the year; we couldn't find any AWS page that says so, so treat those claims as unconfirmed.

Developers have been noticing the gap for a while. On a Hacker News thread about the DeepSeek V4 launch earlier this year, user deaux complained that the newest DeepSeek model available for serverless inference on Bedrock was months old, while antiloper pointed out the trade-off people actually care about: running DeepSeek on AWS keeps data on AWS infrastructure, but costs more than DeepSeek's own pricing (Hacker News).

If you want DeepSeek V4 inside AWS today, SageMaker JumpStart is the official route.

How much does DeepSeek V4 cost on AWS?

There's no per-token price for DeepSeek V4 on AWS. With JumpStart, the model weights are free (MIT licence) and you pay standard Amazon SageMaker AI instance-hour pricing for the endpoint you deploy, for as long as it runs, whether or not anyone sends it a request.

AWS's announcement doesn't name the instance types each model needs, and we couldn't confirm them from an official page, so check the model card in the JumpStart catalog before you deploy. As a rule of thumb, a 1.6T-parameter model like V4-Pro needs a multi-GPU instance with a lot of high-bandwidth memory. V4-Flash, at 284B parameters, is much smaller but is still a multi-GPU job, not a single-card one.

For comparison, here is what DeepSeek charges for its own API, per 1 million tokens, according to its pricing page:

DeepSeek API model Input (cache hit) Input (cache miss) Output
deepseek-flash (DeepSeek-V4.1-Flash-1017) $0.003 off-peak / $0.006 peak $0.15 / $0.30 $0.60 / $1.20
deepseek-v4-pro (DeepSeek-V4-Pro-0813) $0.022 / $0.044 $0.66 / $1.32 $1.98 / $3.96

DeepSeek says off-peak rates are half the peak rates, and that peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. It also says the old deepseek-v4-flash model name is retired and now served by V4.1-Flash at the Flash price.

What that means in practice: for light or bursty use, DeepSeek's API is far cheaper than keeping a GPU endpoint running around the clock. JumpStart makes financial sense when you have steady, high-volume traffic that keeps the instances busy, or when you can't send data to a third-party API at all.

How do you deploy DeepSeek V4 on SageMaker JumpStart?

AWS's announcement points to two routes: the JumpStart catalog in the SageMaker console, or the SageMaker Python SDK. In outline:

  1. Check your GPU quota first. Large JumpStart models need big GPU instances, and new accounts usually have a quota of zero for them. Request an increase in Service Quotas for the SageMaker endpoint instance type listed on the model card. This often takes longer than the deployment itself.
  2. Open SageMaker Studio and go to JumpStart, then search for "DeepSeek".
  3. Pick the model (V4-Pro, V4-Flash or V4.1-Flash) and read its model card: supported instance types, licence and any usage notes.
  4. Choose Deploy, accept the licence terms, pick the instance type and endpoint name, and create the endpoint.
  5. Wait for the endpoint to reach "InService", then call it from the Studio playground, the SageMaker SDK or the InvokeEndpoint API.
  6. Delete the endpoint when you're done testing. Billing runs per instance-hour while it exists, not per request.

Not every deployment goes smoothly. On DeepSeek-V4-Pro's Hugging Face page, a user who tried deploying the model to SageMaker by hand earlier this year reported "facing mulitple issues" [sic], including failed instance health checks and an unknown deployment error, and the thread had no replies when we checked (Hugging Face discussion #199). JumpStart's pre-packaged containers are meant to avoid exactly that kind of DIY trouble, but we haven't seen independent reports on the new JumpStart versions yet.

DeepSeek V4 on AWS vs DeepSeek API vs Bedrock: which should you use?

This is our decision table, based on the official pages listed in Sources:

If you need… Best route Why
The cheapest way to try V4 DeepSeek API Per-token billing from fractions of a cent; no GPUs to manage
V4 with data kept inside your AWS account SageMaker JumpStart Weights run on your own endpoint; no traffic to DeepSeek's servers
Serverless, pay-per-token DeepSeek on AWS Bedrock, but only V3.2 / V3.1 / R1 V4 isn't on Bedrock's serverless list yet
High, steady traffic with a fixed budget JumpStart (do the maths) A busy endpoint can beat per-token pricing; an idle one can't
Multimodal (image) input on AWS V4.1-Flash on JumpStart AWS says it has "native multimodal support"
Fine-tuning or custom serving JumpStart or self-host on EC2/HyperPod You control the weights (MIT licence)
Long documents (up to 1M tokens) Any V4 model All three have a 1M-token context, per AWS

Which DeepSeek V4 model should you pick?

  • V4-Flash is the practical default for most production text work: a much smaller footprint than Pro, the same 1M-token context, and AWS positions it for cost-conscious use.
  • V4.1-Flash is the newest and, per AWS, the most efficient for agents and coding, and the only one of the three AWS describes as multimodal. If you're building an agent on AWS, start here. (For running agents on AWS's managed runtime, see our explainer on AgentCore Runtime V2, which became the default the day after this launch.)
  • V4-Pro is the flagship. Choose it only if you've shown Flash isn't good enough, because its 1.6T parameters make it by far the most expensive to host.

Editor's take

This is a useful launch with a confusing label. Most people searching "DeepSeek V4 AWS" want what Bedrock gives you for Claude, Kimi or GPT models: an API key-style, pay-per-token endpoint with no servers. JumpStart is a different product: you're renting GPUs to host the model yourself, with AWS doing the packaging.

For a regulated company that wants DeepSeek's quality without sending prompts to a Chinese-hosted API, that's a real option, and the MIT licence makes it legally simple. For a hobbyist or a small startup, it's almost certainly more expensive than DeepSeek's API. We'd expect Bedrock to add V4 at some point, since it already hosts three earlier DeepSeek models, but AWS hasn't announced a date, so don't plan around it.

If you're comparing other open-weight and frontier models on AWS, our explainers on Kimi K3.1 Pro on Bedrock and GPT-6.1 Terra on Bedrock cover the per-token alternatives, and our SageMaker JumpStart models update explains how JumpStart works in general.

FAQ

Is DeepSeek V4 available on AWS?

Yes. Since 11 November 2026, DeepSeek-V4-Pro, DeepSeek-V4-Flash and DeepSeek-V4.1-Flash are available on Amazon SageMaker JumpStart, where you deploy them to your own SageMaker endpoint.

Is DeepSeek V4 on Amazon Bedrock?

Not as a serverless model at the time of writing. Bedrock's official model list includes DeepSeek V3.2, DeepSeek-V3.1 and DeepSeek-R1, but no V4 model. AWS hasn't announced a date.

How much does DeepSeek V4 cost on AWS?

There's no per-token price. On SageMaker JumpStart you pay SageMaker AI instance-hour rates for the endpoint you deploy; the model weights themselves are free under the MIT licence.

What is the price of the DeepSeek V4 Pro API?

On DeepSeek's own API, V4-Pro costs $1.32 per million input tokens (cache miss) and $3.96 per million output tokens at peak times, and half that off-peak, according to DeepSeek's pricing page.

What's the difference between DeepSeek V4 Pro and V4 Flash?

V4-Pro is a 1.6T-parameter model with 49B active per token, built for maximum quality. V4-Flash is 284B parameters with 13B active, built for lower cost and latency. Both support a 1M-token context, according to AWS.

Is DeepSeek V4.1 Flash multimodal?

AWS says DeepSeek-V4.1-Flash "provides next-generation agentic efficiency with native multimodal support", including vision tasks, with a 1M-token context and 384K maximum output tokens.

Can I use DeepSeek V4 commercially on AWS?

AWS says the three models are released under the MIT licence, which generally allows commercial use. Read the licence and model card in JumpStart before deploying, and check your own compliance requirements.

Does DeepSeek see my data if I use it on SageMaker?

When you deploy through JumpStart, the model runs on an endpoint in your own AWS account rather than on DeepSeek's API, so your prompts aren't sent to DeepSeek's servers. Your data handling is then governed by your AWS setup.

Sources

Featured image: Server racks inside an AWS data center (a general AWS facility, not specific to DeepSeek or SageMaker). Image: Amazon. Source

Share:

Comments

Leave a comment