deepseek-r1-huoshan-0528

DeepSeek-R1-0528 is the flagship language model launched by DeepSeek, primarily designed to enhance deep thinking and reasoning capabilities through minor version upgrades, delivering a superior user experience.

• Improved Reasoning: Outperforms in benchmark tests such as mathematics and programming, approaching the level of top international models; its accuracy on the AIME 2025 test has improved from 70% to 87.5%.
• Expanded Capabilities: The hallucination rate has been reduced by 45–50%, creative writing has been optimized, tool calls are supported, and multi-domain performance has been significantly enhanced.
• API Optimization: Synchronized updates include support for viewing the thought process, as well as new features like Function Calling and JsonOutput.
• Open-Source Friendly: Adopting the MIT License, an open-source version with a context length of 128K is available.

───────────────────────────────────────────────────────────────────

Core Capabilities

💡 Deep Thinking: After training, more computational resources are devoted, enabling more detailed and thorough reasoning during problem-solving, with an average increase in tokens used per question.

📈 Performance Boost: The distilled chain-of-thought approach has been applied to the 8B model, which excels in mathematical tests and outperforms Qwen3 - 8B.

🌟 Multi-Domain Optimization: Hallucinations have been reduced, creative writing has been refined, and capabilities such as front-end code generation and role-playing have been enhanced.

🛠️ API Enhancements: The meaning of max_tokens has been adjusted, supporting the viewing of the thought process and introducing new features.