DeepSeek unveils V4.1-Flash model with 1m context and enhanced efficiency

DeepSeek AI has released its latest model, DeepSeek-V4.1-Flash, featuring a 1M-token context window, FP4 KV cache, and cross-layer attention reuse. The model, with 552B backbone parameters and 196B additional Engram parameters, is designed to address the challenges of long-horizon agents by reducing the strain on hardware resources. It achieves a global KV cache footprint of 890 bytes per token, significantly smaller than previous versions. The model's architecture includes a 40-layer backbone split into causal encoder and decoder layers, with the decoder leveraging encoder hidden states to minimize prefill compute. The team also introduced CSA2 and hierarchical sparse indexing to optimize memory usage. Deployment benefits include reduced SSD storage and a distributed DRAM pool for SWA KV. The model is available under an MIT license on Hugging Face, with public API tiers for different reasoning levels. Post-training improvements focus on large-scale task synthesis and distillation from over 40 teachers, enhancing efficiency without introducing new algorithms. The model matches DeepSeek-V4-Pro-Base in world knowledge and coding while using fewer parameters. The release follows the earlier announcement of DeepSeek-V4 Pro service, which was delayed to September 14.