DeepSeek releases DeepSeek-V4.1-Flash with 890-byte per-token KV cache
DeepSeek released DeepSeek-V4.1-Flash on Hugging Face. It is a multimodal Mixture-of-Experts model with 552B backbone parameters, 8B activated parameters per token during prefill and 16B during decode, and supports up to 1M token context. The model uses a Causal Encoder-Decoder architecture, Compressed Sparse Attention 2, FP4 main KV caching, and achieves a global KV cache footprint of 890 bytes per token, roughly 1/4 of DeepSeek-V4-Flash. It was trained from scratch on 45T tokens with sparse attention at 64K sequence length and context extended to 1M tokens at 34T tokens. The model is released under the MIT license.
China context
- Original name
- DeepSeek-V4.1-Flash
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can access the model weights on Hugging Face under the MIT license and integrate the model into their own systems.
- For investors
- The release signals DeepSeek's continued focus on efficiency for agentic workloads, which may pressure competitors on inference cost.
DeepSeek released DeepSeek-V4.1-Flash, a multimodal MoE model with 552B parameters and 1M token context, featuring a Causal Encoder-Decoder architecture and Compressed Sparse Attention 2 to reduce KV cache to 890 bytes per token. The model is available on Hugging Face under the MIT license.
The Causal Encoder-Decoder architecture projects the decoder's global KV cache from final encoder hidden states, enabling only 8B parameters per token during prefill and 16B during decode. SWA Bounded Replay reconstructs missing SWA KV states by replaying recent tokens, reducing persistent KV cache to 1/8 of DeepSeek-V4-Flash. CSA2 assigns static attention modes (Full, Reindex, Reuse) and uses a Hierarchical Sparse Indexer to bound deeper indexer cost. FP4 main KV caching with E2M1 format and one E4M3 scale per 16 channels contributes to the 890-byte per-token footprint.
The 890-byte per-token KV cache is a 4-fold reduction from DeepSeek-V4-Flash and 437-fold from DeepSeek-V1, targeting cost efficiency for input-heavy agentic workloads. The MIT license and Hugging Face release enable broad adoption and integration by developers outside China.
The reduced KV cache and activated parameters lower inference costs for long-context and agentic applications, making the model attractive for developers building cost-sensitive AI systems.
Observable next signals include independent benchmarks of DeepSeek-V4.1-Flash against other models on agentic tasks, adoption in open-source frameworks, and API availability or pricing announcements from DeepSeek.