DeepSeek-V4.1-Flash cuts cache demands for million-token contexts
DeepSeek AI has released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 552B backbone parameters, 196B Engram parameters and a 1M-token context window. It activates 8B parameters per token during prefill and 16B during decoding. The reported global KV-cache footprint is 890 bytes per token, about one quarter that of DeepSeek-V4-Flash and roughly 437 times below DeepSeek-V1.
The model uses a 40-layer causal encoder-decoder structure: 20 encoder layers handle prompt processing, while 20 decoder layers generate output. Compressed Sparse Attention 2 shares cache information across layers through Full, Reindex and Reuse modes. The main cache uses E2M1 FP4 quantization with E4M3 scales, while sliding-window attention states are kept in host memory rather than persisted to SSD. The system can replay the last 128 prompt tokens when those states are unavailable.
DeepSeek says the model was trained on 45T multimodal tokens and extended to a 1M-token context using 34T tokens. Open weights are available under the MIT license with vLLM, SGLang and Transformers support. In reported evaluations, the model surpassed Opus-5 and GPT-5.6 Sol on Terminal-Bench 2.1 and DeepSWE v1.1.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.