DeepSeek has released V4.1-Flash, a 552-billion-parameter multimodal model built to lower the cost of long-context inference. It supports contexts up to one million tokens and focuses on shrinking the key-value cache, the stored representation that prevents a model from recomputing earlier context at every step.
According to the technical report, the cache held in GPU memory uses about one-quarter as much space as in DeepSeek-V4-Flash, while the portion offloaded to system memory or storage falls to roughly one-eighth. The model also splits input processing from text generation: it activates 8 billion parameters per input token and 16 billion while producing output. DeepSeek says this nearly halves input-side compute, which matters for agents that repeatedly ingest tool results.
The main cache is stored in lower-precision FP4 format rather than FP8, further reducing its memory footprint. DeepSeek says the model was trained from scratch on 45 trillion text and image tokens.
The company reports competitive coding performance against leading closed models, but also weaknesses on complex scientific work and image analysis. Those comparisons and the cost savings require independent validation under real workloads before operators can judge the trade-offs.