DeepSeek released V4.1-Flash, a new AI model that significantly reduces memory requirements for AI agents by shrinking the KV cache to about a quarter of its predecessor's size. The model uses 552 billion parameters and employs techniques like splitting the architecture into encoder and decoder components to halve compute needs for input processing. Performance matches leading models on coding tasks, though weaknesses remain in scientific reasoning and image analysis.
A researcher replayed real Claude Code and Mooncake requests through a prefix-cache simulator to test whether alternative eviction policies could beat LRU, but failed across three approaches. The analysis reveals that under capacity pressure, most recomputation stems from tool-calling loops seconds apart rather than idle sessions, and the production LRU baseline proves surprisingly difficult to improve upon.
A token leak in Codex's goal mode causes users to burn through $200 weekly subscription limits in a day. The issue stems from the orchestrator restarting the model every 0.03 seconds while waiting for child agents, forcing it to reread 120–470k tokens per continuation until hitting rate limits. Proposed fixes include using models with native sleep support or implementing pauses between continuations.