Cloud Providers Silently Update API Weights, Breaking Coding Agent Performance

An analysis of coding-agent telemetry reveals that cloud providers silently overwrote the deepseek-v4.1-flash model weights on October 19 with version -1017. This update altered tool-calling behavior and degraded overall model performance, leading to a surge in user frustration. This incident highlights the inherent reliability and reproducibility risks of relying on cloud-hosted LLM API endpoints where users cannot control or roll back model updates. It reinforces a key argument for using local open-weight models, which allow developers to freeze weights and maintain stable workflows. The unannounced -1017 update fixed an issue where the model occasionally replied in Chinese, but it made tool calls more aggressively and drastically reduced code quality. Consequently, the user's swearing rate tripled and mid-session model switches increased by 6.7 times.