**[DeepSeek and Xiaomi Join Forces to Usher in the Era of Zero-Cost Context at Scale, Accelerating Agent Scenario Adoption]** According to monitoring by Beating, following Alibaba's Qwen team enabling implicit caching for Qwen3.7-Max to reduce input costs by up to 80%, Xiaomi has announced permanent price reductions for its self-developed MiMo-V2.5 series APIs, aligning their pricing entirely with DeepSeek V4's full series. The flagship model's input cache hit price is $0.0036 per million tokens, cache miss price is $0.435, and output price is $0.87. This alignment aims to intercept global developer traffic and comprehensively accelerate the adoption of intelligent Agent scenarios. Within a month of its April 24 release, DeepSeek V4 Flash topped OpenRouter's monthly leaderboard with a consumption of 7.99 trillion tokens, while V4 Pro entered the top ten. In high-frequency code-reading Agent programming scenarios such as Cursor and Claude Code, thanks to a 99% prefix cache hit rate, developers using the Pro model consumed 80 million tokens for just 4 RMB, while using the Flash model consumed 27.8 billion tokens in a single day for only $160. In comparison, Alibaba's Qwen3.7-Max automatic implicit caching offers only an 80% discount, while explicit caching faces a 125% initial creation premium and a 5-minute lifespan. The high creation premium and short retention period technically imply high system cache construction and maintenance costs, with token-level computational load and KV cache occupancy limiting the scope for cost reductions. Xiaomi and DeepSeek's bold price cuts are made possible by underlying algorithmic advantages. In the inference of 1 million tokens, DeepSeek V4 leverages Compressed Sparse Attention (CSA) and Highly Compressed Attention (HCA) to reduce inference FLOPs to 27% of the previous generation, while KV cache space is reduced to 10%, achieving over a hundredfold reduction compared to traditional GQA models. Xiaomi's MiMo-V2.5-Pro activates only 4.1% (42B) of its total 1.02T parameters, utilizing interleaved Sliding Window Attention (SWA) and Global Attention (GA) to reduce long-context KV cache overhead by sevenfold, while Multi-Token Prediction (MTP) boosts output throughput by threefold. Both solutions squeeze resources at the algorithmic level, heralding the arrival of a low-cost adoption era. [Original Link]