Jefferies has issued a research report noting that DeepSeek's launch of the DeepSeek-V4.1-Flash model represents a highly cost-effective offering in the current AI landscape. The firm has assigned "Buy" ratings to Alibaba-W (09988), Tencent Holdings (00700), and MiniMax-W (00100), with target prices set at HK$184, HK$750, and HK$533, respectively.

The research house points out that the DeepSeek-V4.1-Flash model introduces several key breakthroughs. It features a total parameter count of 552 billion (with active parameters of 8 billion for input and 16 billion for output), adopts a novel Causal Encoder-Decoder (CED) architecture, and utilizes a new pre-training approach complemented by large-scale post-training. In third-party Agentic Benchmark evaluations, the model has outperformed competitors such as Moonshot's Kimi-K3, Anthropic's Opus 5, and OpenAI's GPT-5.6 Sol.

Jefferies also highlights that compared to previous model iterations, the V4.1-Flash has significantly reduced memory requirements, cutting HBM and SSD usage to one-quarter and one-eighth, respectively. DeepSeek continues to employ peak and off-peak pricing strategies. During peak hours, cache hit, input, and output prices are set at RMB 0.04, RMB 2, and RMB 8 per million tokens, respectively. In off-peak periods, these prices drop to RMB 0.02, RMB 1, and RMB 4 per million tokens. When compared at peak-hour rates, these prices are 87%, 78%, and 70% lower than those of the DS-V4-Pro, and also substantially cheaper than rival models: 98%, 90%, and 92% lower than Kimi-K3, and 98%, 75%, and 71% lower than GLM-5.3.

Given the intensifying competitive landscape, the firm anticipates that industry peers will pursue further innovations in their subsequent model iterations. Additionally, Jefferies notes that following the release of the V4.1-Flash model, DeepSeek has adjusted pricing for the older V4 Flash version, reducing peak-hour cache hit, input, and output prices by 60%, 33%, and 11%, respectively, bringing them to RMB 0.04, RMB 2, and RMB 8 per million tokens.