A high-compute model specifically engineered for deep reasoning and complex logical analysis. Leveraging DeepSeek Sparse Attention, it excels in mathematical, algorithmic, and research-grade tasks. Optimized through extensive reinforcement learning, it offers exceptional performance on long-context scenarios, making it an ideal choice for academic exploration and solving highly complex problems.

A groundbreaking multimodal and multilingual reranker designed for visual document retrieval. It supports 29+ languages and seamlessly processes text, images, and mixed content. By unifying modalities to eliminate the "modality gap," it significantly enhances search precision and relevance for RAG systems in long-document, complex-layout, and code-retrieval scenarios.

Kimi-K2.5-CC is a general-purpose language model from Moonshot AI designed for text generation and conversational applications. It focuses on practical instruction following, knowledge understanding, and content creation. Core capabilities include multi-turn dialogue, summarization, question answering, rewriting, and common writing assistance. It is suitable for AI assistants, content production, enterprise knowledge Q&A, and everyday office automation workflows.

As Anthropic's most capable Sonnet model to date, this version delivers a full upgrade across coding, computer use, and agentic planning. With advanced reasoning capabilities and a 1M token context window, it excels at long-document analysis, complex codebase maintenance, and automated workflows, making it the ideal choice for balancing high performance with productivity.

Built upon advanced chain-of-thought and sparse attention mechanisms, this model is optimized for complex reasoning and agentic tasks. Through large-scale reinforcement learning, it achieves deep logical deduction capabilities with support for efficient tool calling and structured output. It offers a superior balance of high performance and low latency in mathematics, programming, logical reasoning, and long-context agentic workflows.

A reasoning-enhanced large language model featuring exceptional capabilities in code generation, mathematical computation, and logical deduction. It supports a 256K long context window and integrates native multimodal understanding with advanced chain-of-thought reasoning. Designed for complex analysis and high-stakes tasks, it excels in web search and tool utilization, providing high-precision intelligence for various professional scenarios.

Designed for high-performance efficiency, this reasoning model leverages large-scale reinforcement learning to optimize internal deliberation, offering superior intelligence density with significantly reduced latency. It features a massive context window, integrated web search, and robust tool-use capabilities. Perfectly balancing cost-effectiveness with frontier-level intelligence, it excels at complex analytical tasks and multimodal data processing.

As the most efficient and lightweight model in the GPT-5.4 series, it is specifically designed for high-concurrency, low-latency production environments. While maintaining high cost-effectiveness, it excels in data extraction, classification, ranking, and lightweight coding tasks. It serves as an ideal execution engine for building large-scale agentic systems and automated pipelines, delivering fast and reliable responses at minimal cost.

A high-performance, cost-effective hybrid reasoning model based on the MoE architecture. It supports web browsing, tool calling, and complex code generation. Featuring flexible reasoning modes, it dynamically adjusts its thinking depth based on task complexity, ensuring both rapid responses and advanced logic for building intelligent agent applications.

A highly efficient and lightweight image-to-video generation model that creates high-coherence, cinematic video clips of 5 to 10 seconds from a single reference image or start/end frames combined with text prompts. It maintains strong visual consistency and offers precise camera motion control, making it ideal for e-commerce marketing, dynamic posters, and rapid creative content production.