DeepSeek officially released its V4.1 Flash model on September 10 — a 552-billion-parameter multimodal mixture-of-experts architecture with a 1-million-token context window, which the company claims beats V4 Pro on performance, cost, and speed.
The model adds native multimodal visual understanding, something earlier Flash versions didn't offer in an integrated way. On the architecture side, it uses an FP4 KV cache combined with cross-layer attention reuse — a combination DeepSeek says shrinks the overall KV cache to just 890 bytes per token, the key to sustaining a 1M-token context window without memory costs exploding.
In the API, the model is now called deepseek-flash, and old references to v4-flash are automatically redirected to this new version, so existing integrators don't need to change any code.
The timing of the launch hasn't gone unnoticed: it lands just two days after the NSA, CISA, and the FBI formally accused DeepSeek of distilling U.S. frontier models — including GPT-5, Gemini 2.5 Pro, and several versions of Claude — to train precisely the R1 and V3 model line that V4.1 directly descends from. DeepSeek had not publicly responded to the alert as of this writing.
For technical teams running agents or high-volume RAG pipelines, a Flash-tier model with a 1M-token context and a drastically lighter KV cache translates into lower inference cost per long document processed. But adopting open weights from a Chinese model for production workloads means self-hosting, reviewing data residency, and now also weighing the compliance risk flagged in the U.S. agencies' alert.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Get in touch for a consultation.
Contact Carlos Montiel