DeepSeek-V4.1-Flash has recently been launched on the Qwen AI platform, with API services and Token Plan now available. Developers can integrate the model into their systems via standard APIs, or directly use the Token Plan in tools such as Qoder, Qwen APP, and Codex for code, documentation, visual understanding, and agent tasks.
According to the platform announcement, this model is DeepSeek's new lightweight flagship, featuring a MoE architecture with 552B total parameters, using a Causal-Encoder-Decoder asymmetric structure, with input activation of about 8B and output activation of about 16B; it natively supports text and image understanding, with a maximum context length of 1 million tokens, and a maximum output of approximately 393K. The official stated that it has improved significantly over its predecessor in several Agent and code benchmarks, achieving high throughput and low latency with lower activation parameters.
The cost side is a major focus of this release. The new generation of cache compression reduces the KV Cache demand for HBM to one-quarter of the previous generation and for SSD storage to one-eighth, saving more resources for long contexts and multi-turn tool calls. The Qwen page provides time-based pricing: 1 yuan per million tokens during off-peak hours for input and 4 yuan per million tokens for output; 2 yuan and 8 yuan respectively during peak hours. The rate limits are 15K RPM and 1M TPM, covering features such as prefix completion, function calls, caching, structured output, batch tasks, and online search.
In terms of deployment, Alibaba Cloud BaiLian has collaborated with the vLLM open-source community to complete the adaptation, making it accessible on both the China site and international site, covering regions including Beijing, Singapore, and Global locations such as the United States, Germany, Japan, and Hong Kong. According to BaiLian documentation, the model supports multi-turn conversations, function calls, online search, context caching, and structured output, with a max_tokens limit of approximately 393,216, making it suitable for enterprises to migrate tasks such as long document parsing, customer service knowledge base, code repository Q&A, and multimodal order review to a single interface.
Industry experts believe that by incorporating the combination of "large parameters, small activation, strong caching" into the Qwen ecosystem, V4.1-Flash will reduce the trial and error costs of long-context agents; for small teams, the low off-peak price combined with the Token Plan subscription allows for more flexible batch inference and prototype validation.
