DeepSeek V4 Preview: What the Fast, Expert, and Vision Modes Suggest

Community signals and unverified artifacts suggest DeepSeek V4 may be in development, featuring three distinct inference modes: Fast (optimized for speed and cost), Expert (deep chain-of-thought reasoning), and Vision (multimodal image understanding). Fast Mode targets latency-sensitive, high-volume use cases at low cost, competing with GPT-4o-mini and Claude Haiku. Expert Mode extends R1-class reasoning within a unified API, giving developers deterministic control over quality and cost. Vision Mode integrates image comprehension as a first-class capability, potentially offering a self-hosted alternative to GPT-4o and Gemini for visual workflows. The mode-based architecture lets developers explicitly choose their performance-cost tradeoff at call time rather than relying on opaque routing. Key unknowns include pricing per mode, context window sizes, fine-tuning availability, and whether open weights will ship for all three modes. A practical pre-release checklist covers API monitoring, mode-mapping existing features, building pricing scenarios, preparing benchmark prompts, and compliance review for Chinese-origin model policies.