Chinese tech media reported an undisclosed DeepSeek model update on March 9 with expanded context handling, informally labeled 'V4 Lite.' Circulating internal benchmark claims include 90% HumanEval Pass@1 and SWE-bench above 80%, which would match Claude Opus 4.6 and GPT-5.3 Codex on coding tasks. No third-party verification exists as of March 15, and DeepSeek has not officially confirmed the update — treat as single-source rumor.