Ant Ling has released Ling-3.2-flash, a natively multimodal model featuring 560B total parameters, 25B activated per step, and support for a 1M context window. The model introduces visual feedback self-checking and anomaly-adaptive recovery, enabling it to use rendering and interface outputs as observational input for self-correction, or to locate and retry after errors or tool failures. It employs a 7:1 hybrid attention scheme consisting of seven KDA layers and one Gated MLA layer. With DSpark speculative inference, it achieves an average single-user decoding speed of 280 tokens/s for a long context of approximately 30K tokens.