DeepSeek-V4 preview drops with 1.6T parameters, 1M context

DeepSeek-V4 is a strategic strike against the high inference costs of Western frontier models, offering GPT-5 level reasoning at an order of magnitude lower price point.

• –The 1.6T parameter MoE architecture is highly efficient, activating only 37-49B parameters per token to enable low-latency inference on high-end consumer hardware.
• –Native multimodality (text, image, video, audio) trained from scratch indicates DeepSeek is no longer just "catching up" but setting the pace for integrated foundation models.
• –"Engram Memory" and "mHC" architectures solve the long-context degradation problem, making it a premier choice for repository-level coding tasks.
• –By maintaining an aggressive Apache 2.0 open-weights stance, DeepSeek is effectively commoditizing high-end intelligence and forcing a pricing war.
• –Optimization for "Thinking Mode" with adjustable reasoning effort makes it the most flexible backbone for autonomous AI agents currently on the market.