Amazon adds 3 open DeepSeek V4 models to SageMaker JumpStart
The model DeepSeek-V4-Pro has 1.6 trillion parameters, but uses only 49 billion for each token.
Amazon tells that 3 models from DeepSeek are in Amazon SageMaker JumpStart. It is a catalog of models that a team can start in a small number of clicks. The models have open weights (the model files are open to all) and the MIT license. Amazon tells that they are frontier-class models but gives no test results.
How it works
All 3 are Mixture-of-Experts models. Each token (a small piece of text) uses only a part of the model. All 3 hold a context window of 1 million tokens.
- V4-Pro: 1.6 trillion parameters, 49 billion active for each token.
- V4-Flash: 284 billion parameters, 13 billion active, for a lower cost.
- V4.1-Flash: 552 billion parameters. It also reads images.
The numbers
Amazon tells that V4.1-Flash cuts its cache (memory for the text that came before) to one quarter of the cache of V4-Flash. Amazon gives no speed, cost or precision results for the models.
What is not known
The source gives no price. It does not tell how the models compare with other models in a test. A team must test them on its own tasks.
Insight: AWS sells the compute for free models
Amazon added the Nemotron 3.5 Super VL model of NVIDIA on 9 November. On 22 October it opened its training tool to 4 open models. This reads as a move to make AWS the primary host of open models: the model is free, but the hosting is not. Leaders get more alternatives, and more lock-in to one host.
Sources
Posted
