DeepSeek-V4-Flash-0731 is the official release of V4-Flash, featuring a massive leap in agentic capabilities. It outperforms V4-Pro (Preview) on key benchmarks, natively supports the Responses API, and is fully adapted for Codex CLI.
DeepSeek-V4 Preview is a new series of highly efficient MoE language models, featuring V4-Pro (1.6T params) and V4-Flash (284B params). Both models support a 1 million token context window by default, utilizing a novel hybrid attention architecture to drastically reduce compute and memory costs.
DeepSeek-OCR is a model that compresses long text by treating it as an image. This optical compression uses far fewer vision tokens to represent documents, unlocking new levels of efficiency for long-context tasks while delivering powerful OCR capabilities.
DeepSeek V3.2 and V3.2-Speciale are breakthrough open-source AI models from China, built to rival top closed models like OpenAI’s GPT-5 and Google Gemini 3 Pro. Delivering gold-level results in both math and programming benchmarks, DeepSeek V3.2-Speciale even outperforms GPT-5 on AIME 2025. With efficient architecture and exceptional reasoning abilities, these models are perfect for both enthusiasts and enterprise applications
Powered by the groundbreaking DeepSeek-V3 model with over 600B parameters, this state-of-the-art AI leads global standards and matches top-tier international models across multiple benchmarks.
DeepSeek-V3.1-Terminus is the latest update to the DeepSeek-V3.1 model. This "Terminus" version focuses on stability and refinement, fixing issues like language mixing and improving agent capabilities, while retaining the core strengths of the V3.1 series.
DeepSeek-V3.2-Exp is a new experimental model introducing DeepSeek Sparse Attention (DSA). This new architecture boosts long-context efficiency for training and inference while maintaining the performance of V3.1-Terminus. API prices have been cut by over 50%.
DeepSeekMath-V2 is a new open-source model specialized in mathematical reasoning. It introduces a self-verification mechanism where the model acts as both generator and verifier to refine its own proofs. It achieved Gold-level scores in IMO 2025 and a near-perfect 118/120 in Putnam 2024.
