DeepSeek官宣！新模型、新突破、新价格

Core Insights - DeepSeek officially released DeepSeek-V3.1, a large model featuring a hybrid reasoning architecture that supports both thinking and non-thinking modes, resulting in higher efficiency compared to its predecessor DeepSeek-R1-0528 [1] - The new model shows significant improvements in tool usage and agent tasks, with notable advancements in code repair assessments and complex task testing in command-line environments [1] - The release is seen as a step towards the "Agent era" in AI, with market predictions estimating the Chinese AI agent market to reach 6.9 billion yuan by 2025 and nearly 30 billion yuan by 2030 [1] Performance Enhancements - Testing results indicate that DeepSeek-V3.1's efficiency in thinking mode has improved significantly, achieving similar average performance to R1-0528 while reducing output token count by 20%-50% [2] - In non-thinking mode, the output length has been effectively controlled, which helps users manage costs [2] API Pricing Adjustments - Starting from September 6, 2023, DeepSeek will adjust its API pricing, removing the previous night-time discount. The new pricing will be 0.5 yuan per million input tokens (cache hit) / 4 yuan (cache miss), and 12 yuan per million output tokens [2] - The previous API pricing was 0.5 yuan per million input tokens (cache hit) / 2 yuan (cache miss), and 8 yuan per million output tokens [2] Technical Specifications - DeepSeek-V3.1 utilizes UE8M0 FP8 Scale parameter precision, which is designed for the upcoming generation of domestic chips [2] - Recent tests by the China Academy of Information and Communications Technology indicate that products deploying the DeepSeek model have achieved accuracy in language understanding and logical reasoning tasks comparable to foreign systems [3]