DeepSeek: Shrinks Cache 437x for AI Agents

DeepSeek trimmed agent memory to 890 bytes of cache per token, delivering a 437x reduction versus DeepSeek-V1 while adding just 25% compute per output token as context scales from 4K to 1M. The resulting Flash model posted 39 on the AA Index against V4-Pro’s 36 and cut cost per task to $0.27 from $0.67, underscoring how LLM optimization techniques and AI agents performance gains now hinge on aggressive KV-cache compression rather than raw parameter counts. Model efficiency gains like these directly improve long-context agent workflows that dominate current production deployments.