DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash models now available on Amazon SageMaker JumpStart

Amazon SageMaker JumpStart has added three DeepSeek models—DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash—expanding the range of foundation models available to AWS customers. These open-weight Mixture-of-Experts (MoE) and Causal Encoder-Decoder (CED) models are released under the MIT license and target reasoning, coding, and agentic workloads. The new additions provide customers with options to balance performance, cost, and efficiency based on their specific needs.

DeepSeek-V4-Pro is a 1.6T-parameter MoE model designed for complex reasoning and high-quality text generation. It activates 49B parameters per token and supports a 1M-token context window, making it suitable for long-document analysis and multi-step reasoning tasks. The model is positioned for flagship-scale applications requiring maximum capacity and precision in generation.

DeepSeek-V4-Flash is a 284B-parameter MoE model optimized for lower compute costs while maintaining strong performance. It activates 13B parameters per token and also supports a 1M-token context window, ideal for latency-sensitive and cost-conscious production deployments across diverse language tasks.

DeepSeek-V4.1-Flash introduces a 552B-parameter CED architecture with native multimodal support. It activates 8B parameters during prefill and 16B during decode, supporting a 1M-token context window and 384K max output tokens. Its Compressed Sparse Attention 2 and FP4 KV caching reduce the global KV cache footprint to roughly 1/4 of V4-Flash, improving cost efficiency for agentic workloads, coding, and vision tasks.