DeepSeek AI has released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model featuring a 1-million-token context window and an FP4 KV cache that reduces global cache footprint to 890 bytes per token. The model utilizes a causal encoder-decoder architecture and Compressed Sparse Attention 2 (CSA2) to significantly lower memory requirements while maintaining performance.

  • The model has 552B backbone parameters, 196B Engram parameters, and activates 8B parameters during prefill and 16B during decode.
  • A causal encoder-decoder structure halves prefill compute by deriving the decoder's global KV from the encoder's final hidden state.
  • CSA2 assigns layers to Full, Reindex, or Reuse modes to share main KV caches and indexer indices across layers.
  • FP4 quantization of the main KV cache cuts storage against V4's FP8 cache by nearly half.
  • The model matches DeepSeek-V4-Pro-Base on world knowledge and coding while using 1/3 of total parameters.
  • It outperforms Opus-5 and GPT-5.6 Sol on Terminal-Bench 2.1 and DeepSWE v1.1 benchmarks.

The open-weight model is available under an MIT license via Hugging Face with support for vLLM, SGLang, and Transformers, offering a public API with low, high, and max reasoning tiers.