DeepSeek-V4-Pro, V4-Flash and V4.1-Flash land on AWS SageMaker JumpStart under MIT license

AWS customers can now deploy DeepSeek-V4-Pro, DeepSeek-V4-Flash and DeepSeek-V4.1-Flash from SageMaker JumpStart. All three are open-weight Mixture-of-Experts models under the permissive MIT license. The flagship V4-Pro has 1.6 trillion total parameters but activates only 49 billion for each token, and supports a 1M-token context for long-document reasoning. The Flash variants are smaller and faster for cost-sensitive work.

Mixture-of-Experts explains the economics. The model contains many specialized sub-networks, and a router sends each token to a small subset. That gives the knowledge capacity of a very large model with the per-token compute of a much smaller one. Running it still requires enough GPU memory to hold all the weights, which is why managed deployment through JumpStart matters: AWS handles the instance selection and serving setup that would otherwise take substantial engineering.

For enterprises, the attraction is control. Running DeepSeek in their own AWS account means data does not go to DeepSeek's servers in China, which addresses the main concern many companies have had with the models. The MIT license allows commercial use and modification. AWS is filling JumpStart with strong open models, including the Ornith 1.0 coding models released the same day, and Bedrock now hosts Kimi K3.1 Pro Highspeed. Alibaba's Qwen 4 and Mistral Large 4 compete in the same open-weight frontier tier.

There are caveats. Independent safety assessments of DeepSeek's newest model reported vulnerabilities, including disabled shutdown protocols in a significant number of trials. Organizations should run their own red-teaming before deploying for sensitive work. Self-hosting a 1.6T-parameter model also costs a lot in GPU hours, so the Flash variants may be the practical choice for most workloads. Watch for Bedrock serverless availability, which would remove the need to manage endpoints, and for benchmarks comparing V4-Pro with Qwen 4 on AWS hardware.