DeepSeek Releases V4.1-Flash: Replaces V4-Pro and Optimizes KV Cache

On September 10, 2026, DeepSeek released V4.1-Flash, a 552B parameter MoE multimodal model. This model replaces the previous flagship V4-Pro and significantly improves the cost structure by reducing the KV cache to 890 bytes per token. This is 1/4 of V4-Flash and 1/437 of V1.

Architecture and Technical Innovations

• Causal Encoder-Decoder (CED): Reduces prefill computation to O(NL/2) and improves efficiency by separating the encoder and decoder.
• CSA2 Compression: Performs simultaneous compression across three axes—items, sequences, and layers—while limiting the block candidate pool through a hierarchical sparse indexer.
• SWA and FP4 KV: Minimizes memory usage by applying Bounded Replay, which reuses only the last 128 tokens, and NVFP4-based quantization.
• Engram and DSpark: Enhances inference speed through the 196B parameter Engram module and the DSpark draft generator based on a 3-block sliding window.

Performance and Benchmark Results

• Computational Efficiency: Even when context length increases 256-fold from 4K to 1M, decode FLOPs increase by only 1/4.
• Benchmarks: Claims to surpass Opus 5 and GPT-5.6 Sol with scores such as 90.6% on TB2.1, 74.2% on DeepSWE, and 3471 on Codeforces. However, it shows inferior performance in high-difficulty reasoning or HLE without tool usage.
• Multilingual Performance: Korean token efficiency is approximately 1.4 times less efficient than Chinese, with weaknesses in multilingual performance shown by scores of 45.5 on MultiLoKo and 80.2 on MGSM.

Deployment and Infrastructure

• API and Cost: The deepseek-flash API is 7.3 times cheaper for input and 3.3 times cheaper for output on cache hits compared to V4-Pro. Starting September 14, 2026, V4-Pro will be routed to V4.1-Flash.
• Open Source: Weights are released on Hugging Face under the MIT license, allowing free commercial use and modification.
• Limitations: The simplified architecture may lead to capability degradation on extreme inputs, and close benchmark scores do not imply full parity with frontier models.