DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

DeepSeek releases V4.1-Flash, a 552B-parameter multimodal MoE with 1M-token context whose CSA2 + FP4 caching compresses runtime KV cache to 890 bytes/token (~1/4 of DeepSeek-V4-Flash) and persistent cache to ~1/8, targeting long-horizon agentic workloads;…

DeepSeek released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 552B backbone parameters and contexts up to one million tokens. Its Causal Encoder-Decoder (CED) architecture activates 8B parameters per token during prefill and 16B during decode. Joint optimization of architecture, precision, and deployment combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching, reducing the runtime/global KV cache to 890 bytes per token — about one-quarter of DeepSeek-V4-Flash (described by one report as 4x KV cache compression) — while SWA Bounded Replay cuts the persistent SSD/host-memory cache to roughly one-eighth at the same sequence length. Pretrained on 45T multimodal tokens, the model reportedly achieves about 420 tokens/s and outperforms the baseline across text and multimodal agentic scenarios, targeting input-heavy, long-horizon agent serving where storage and bandwidth are bottlenecks. Checkpoints are released on Hugging Face under deepseek-ai. According to the Hacker News analysis author, all DeepSeek-V4 Pro models were taken offline following the release.

• 552B-parameter multimodal Mixture-of-Experts model supporting contexts up to 1 million tokens.
• CED (Causal Encoder-Decoder) architecture activates 8B parameters per token during prefill and 16B during decode.
• CSA2 cross-layer KV reuse plus FP4 KV caching reduce the runtime/global KV cache to 890 bytes per token, about one-quarter (4x compression) of DeepSeek-V4-Flash.
• SWA Bounded Replay cuts the persistent SSD/host-memory KV cache to roughly one-eighth of DeepSeek-V4-Flash at the same sequence length.
• Pretrained on 45T multimodal tokens; reported throughput of about 420 tokens/s.
• Checkpoints released on Hugging Face under the deepseek-ai organization.