DeepSeek has recently launched its new-generation AI model V4.1-Flash. According to the company's official introduction, although V4.1-Flash's total and active parameter counts are both lower than those of V4-Pro, it delivers stronger performance, inference speed and cost control, particularly excelling in agent-related and programming tasks, with some benchmark results surpassing V4-Pro and frontier models such as GPT-5.6 Sol.
DeepSeek V4.1-Flash delivers strong performance and cost efficiency
DeepSeek notes that V4.1-Flash uses a mixture-of-experts architecture, with a total of 552 billion parameters, 8 billion active during input and 16 billion active during output. By comparison, the previous-generation V4-Flash has 284 billion total parameters, while V4-Pro reaches 1.6 trillion. A major improvement in the new model is a substantial compression of the KV cache, bringing cache usage down to roughly 890 bytes per token, while output throughput reaches about 214.4 tokens per second and time to first token is 1.13 seconds.
Independent evaluation firm AA described it as "significantly faster."
On performance, DeepSeek says V4.1-Flash, under high reasoning intensity settings, leads GPT 5.6 Sol and Claude Opus 5 in tests covering terminal operations, code repair and automation tasks. However, on broader, more difficult general reasoning tasks, the model's results are comparatively modest. AA's data show V4.1-Flash scores a composite intelligence index of 40, above the median of most comparable open-weight models but still below leading frontier models.
It is worth noting that AA recently upgraded its evaluation framework, tightening overall scoring standards noticeably, and several previously high-scoring models have seen their scores reduced.
On pricing, AA's data show V4.1-Flash is priced at US$0.30 (about HK$2) per 1 million input tokens and US$1.20 (about HK$9) per 1 million output tokens, for a blended cost of roughly US$0.18 (about HK$1), below the median for models in its tier. With cache hits, pricing can drop further to US$0.01 (about HK$0) per 1 million tokens. AA also notes that the model tends to produce longer answers, with total output token volume in evaluations markedly higher than comparable models, which may affect time and cost in real-world use.
Technically, V4.1-Flash introduces a new structure called CED, which splits 40 Transformer layers into 20 causal encoders and 20 decoders, reducing the redundant computation caused by traditional cross-attention. It combines this with methods such as CSA2 and FP4 quantization to compress cache and memory usage. Overall, this update focuses on optimizing agent execution, coding capability and deployment cost.