DeepSeek Releases DeepSeek V4.1 Flash Multimodal MoE Model
DeepSeek has officially launched DeepSeek V4.1 Flash, a 552-billion parameter multimodal Mixture-of-Experts (MoE) model built on an asymmetric Causal-Encoder-Decoder architecture. The model activates only 8 billion parameters on input processing and 16 billion parameters during output generation, achieving high intelligence at low computational cost. By drastically compressing the KV cache—lowering HBM demands to one-fourth and SSD needs to one-eighth compared to the previous generation—the model significantly reduces billing costs for long-context and agent-style tasks. Benchmark results indicate that V4.1 Flash outperforms the older flagship V4 Pro, leading DeepSeek to reduce API pricing and plan an orderly phase-out of V4 Pro. The open-source weights and a technical report are available on Hugging Face, though large-scale self-hosted deployment requires extensive hardware such as a 2k-GPU cluster. DeepSeek has also updated its API with peak and off-peak pricing structures, with off-peak calls priced at half rate.
## BACKGROUND
Mixture-of-Experts (MoE) is an architecture that routes inputs to specialized sub-networks, allowing language models to scale parameter counts without proportionally increasing compute costs for each token. Key-Value (KV) caching is an inference technique that stores previously generated token representations in memory to speed up output, but long-context processing can consume enormous amounts of high-bandwidth memory (HBM).