DeepSeek-V4 preview lands with 1M context

This is a serious price-performance shot across the bow, not just another model drop. DeepSeek is betting that open weights, long context, and strong agentic behavior matter more than multimodal flash than pure benchmark theater.

• –The 1M context default is the headline feature: it makes DeepSeek much more practical for long codebases, document-heavy workflows, and agent loops
• –Splitting the line into V4-Pro and V4-Flash is a smart segmentation move, giving teams a quality tier and a cheaper latency-focused tier
• –Official docs emphasize agentic coding and compatibility with common APIs, which should speed adoption in developer stacks
• –The release is notable for being text-only; that keeps the scope focused, but leaves room for rivals with multimodal depth to differentiate
• –If the reported benchmark claims hold up, V4 will pressure closed-model pricing even where it does not fully beat the top proprietary systems