According to the Guojin Securities Research Report, DeepSeek V4.1 Flash was released on September 10, with native multi-mode, 1 million token contexts, 384K output; asymmetric activation, the cache was reduced to 890 bytes/token, and the video memory was reduced to 1/4 of the previous generation. The API price dropped drastically, but the weight was 510GB, and the local 8-card node was required, and the TCO increased after deployment. Of the seven tests, the code and five agents surpassed the V4 Pro, and the two inference alone fell short of two. Financial Investment Research Institute: The draft valuation can be recalculated, and the industry research development page is second only to GPT-6 Astra. The backtest is strict but lacks transaction restrictions. Focus on large domestic models and AI applications.

According to the Guojin Securities Research Report, DeepSeek V4.1 Flash was released on September 10, with native multi-mode, 1 million token contexts, 384K output; asymmetric activation, the cache was reduced to 890 bytes/token, and the video memory was reduced to 1/4 of the previous generation. The API price dropped drastically, but the weight was 510GB, and the local 8-card node was required, and the TCO increased after deployment. Of the seven tests, the code and five agents surpassed the V4 Pro, which was less than two of the deductions. Financial Investment Research Institute: The draft valuation can be recalculated, and the industry research development page is second only to GPT-6 Astra. The backtest is strict but lacks transaction restrictions. Focus on large domestic models and AI applications.