deepseek-v4.1-flash-1017: Oct 17 iterative MoE multimodal model. 1M context window, up to 384K output, native vision support. RL-optimized for coding & Agent tasks, low latency and high throughput.