Chinese AI firm DeepSeek has launched a preview of its next-generation DeepSeek-V4 model, offering open-weight access via Hugging Face and API services. The Expert (Pro) version boasts 1.6 trillion total parameters (49 billion active) for advanced math, STEM, and coding tasks, while the Instant (Flash) model (284 billion total parameters, 13 billion active) targets simpler agentic tasks. Both support a 1 million token context window, enabling extended reasoning.

The Pro model matches closed-source rivals like Google’s Gemini-3.1-Pro in world knowledge, per DeepSeek’s claims, though it trails slightly in raw performance. The Flash variant surprisingly rivals Pro’s reasoning on basic tasks despite smaller active parameters. Open weights allow community optimization, though running these models locally demands high-end hardware. API pricing starts at $0.01 per 1,000 tokens for Flash and $0.02 for Pro, undercutting competitors.

DeepSeek-V4 integrates with tools like Claude Code and OpenClaw, bridging gaps between research and practical applications. While consumer hardware struggles with the base models, quantized versions could democratize access. The company emphasizes Pro’s dominance in technical domains, positioning it as a open alternative to proprietary systems.

DeepSeek-V4’s open-weights approach accelerates innovation but raises questions about hardware accessibility. By balancing API affordability and model transparency, the firm challenges closed ecosystems while acknowledging current computational barriers.