On September 10, DeepSeek V4.1 Flash made its official debut. This latest iteration boasts native multimodal capabilities, enabling it to support a context window of up to 1 million Tokens and an output capacity of 384K. Leveraging asymmetric activation technology, it dramatically cuts down on cache and memory demands, while simultaneously slashing API pricing. Nevertheless, its substantial weight of 510GB necessitates an 8-card node for local self-deployment, resulting in a higher total cost of ownership (TCO). In a series of seven tests, DeepSeek V4.1 Flash outshone its predecessor, V4 Pro, in five categories, including code generation and agent tasks, with only a slight lag in two reasoning tasks. Within the realm of financial investment research, the model excels at recalculating valuation drafts, and its prowess in constructing entity pages for industry research places it second only to GPT-6 Astra. Although its backtesting criteria are stringent, they fall short in terms of trading constraints.