The lightweight version of DeepSeek V4, codenamed "sealion-lite," is currently undergoing internal testing with at least one inference service provider, with relevant parties having signed non-disclosure agreements. This version supports a 1,000,000-token context window, natively incorporates multimodal capabilities, and outperforms existing web and application-end models in terms of performance. Concurrently, some users have showcased content generated by V4.
