DeepSeek-V4 Preview: Million-Token Context & Agent Upgrades

DeepSeek has launched and open-sourced its new model series, DeepSeek-V4, featuring two models: the flagship DeepSeek-V4-Pro, which uses a Mixture of Experts (MoE) architecture with 1.6 trillion parameters and 1 million token context, and the more cost-efficient DeepSeek-V4-Flash, which is smaller and faster. Both models are designed to enhance agentic capabilities, world knowledge, and reasoning, with V4-Pro showing substantial improvements over previous models and even rivaling top closed-source models in these areas. The introduction of a novel attention mechanism and DeepSeek Sparse Attention in these models achieves long-context performance while reducing computational demands. DeepSeek-V4 is particularly optimized for agent products like Claude Code and OpenCode, ensuring consistency in structured agent tasks. The models are accessible via API and offer both thinking and non-thinking modes, with a deprecation notice for older model names. The launch underscores DeepSeek's commitment to a 1M token context standard and its focus on agent-first optimization, positioning these models as serious contenders in open-source AI for complex reasoning and long-document processing. The models will also be available on the Atlas Cloud platform, which provides production-grade AI access with a focus on long-context workloads and agent pipelines.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.