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

arXiv: https://arxiv.org/abs/2609.19969 · HTML: https://mlsanity.com/paper/2609.19969/

## Key facts

- Cited by 76 papers in the map (data through 14 Oct 2026).
- Placed in Efficient MoE training and inference systems, one of 12 directions in Mixture-of-experts.
- First posted on arXiv 17 Sep 2026.

## Placed in

- Mixture-of-experts / [Efficient MoE training and inference systems](https://mlsanity.com/moe/efficient-moe-training-and-inference-systems/)
- Efficient LLM inference and serving / [KV cache optimization](https://mlsanity.com/inference/kv-cache-optimization/)

## Built on by (most cited first, 30 of 76)

- [HySparse2: Hybrid Sparse Attention with Two-Level KV Sharing](https://mlsanity.com/paper/2609.26368/) (2026)
- [RSI-Router: Evolving Subtask-Level LLM Routing and Skills for Cost-Efficient Agents](https://mlsanity.com/paper/2609.34712/) (2026)
- [Code4Scene: Benchmarking Coding Agents for Constructing and Editing 3D Scenes](https://mlsanity.com/paper/2609.36777/) (2026)
- [KV-Kaizen: Learning Context-Adaptive Cache Compression Choices](https://mlsanity.com/paper/2609.37988/) (2026)
- [Adapting Tree-Structured Speculative Decoding to DeepSeek-V4 for Efficient Inference](https://mlsanity.com/paper/2609.24698/) (2026)
- [KITE: KV-Invariant Transformer Expansion for Efficient Agentic LLM Scaling](https://mlsanity.com/paper/2609.27294/) (2026)
- [Enabling Timely Guidance before Skill Retrieval: Retaining Helpful Warm Tips in Agent Context](https://mlsanity.com/paper/2609.32339/) (2026)
- [VPEvolve: A Self-Evolving Virtual Process Engineer for Computational Lithography](https://mlsanity.com/paper/2609.32473/) (2026)
- [SoFT: Soft Targets for Generalizable LLM Fine-Tuning](https://mlsanity.com/paper/2609.32493/) (2026)
- [ECG-Scroll: A Long-Horizon, Streaming Benchmark and Agent Environment for Interpretation of Ambulatory Electrocardiograms](https://mlsanity.com/paper/2609.33117/) (2026)
- [GSM: Efficient Language Modeling with Shared Global State](https://mlsanity.com/paper/2609.33465/) (2026)
- [A Cheap Verifier is Good Enough: LLM Post-training is Robust to Erroneous Rewards](https://mlsanity.com/paper/2609.33467/) (2026)
- [When Do Agents Help? Embedding, LLM and Agentic Alignment of Classical Texts and Their Translations](https://mlsanity.com/paper/2609.33691/) (2026)
- [Understanding Confabulation and Rethinking Reconstruction in Activation Explanations](https://mlsanity.com/paper/2609.33702/) (2026)
- [Org-Agent: Beyond Personal Assistants Towards Organizational Agents](https://mlsanity.com/paper/2609.34392/) (2026)
- [OmniTide: Co-Designing Algorithms and Systems for Efficient On-Device Omni-LLM Streaming](https://mlsanity.com/paper/2609.34653/) (2026)
- [VCN-Bench: A Video-Contextualized Navigation Benchmark for Spatial Reasoning over Prior Visual Experience](https://mlsanity.com/paper/2609.34687/) (2026)
- [Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation](https://mlsanity.com/paper/2609.35347/) (2026)
- [FactorEngram: Factorized N-gram Memory with Basis-Level Gating for Language Models](https://mlsanity.com/paper/2609.35578/) (2026)
- [MMSkillRisk: Can Agents Stay Safe When Multimodal Skills Become Traps?](https://mlsanity.com/paper/2609.35912/) (2026)
- [Lost in Translation: Measuring the Effect of Non-Native English on End User Performance of Large Language Models](https://mlsanity.com/paper/2609.36214/) (2026)
- [Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression](https://mlsanity.com/paper/2609.36322/) (2026)
- [MOBA-VL: Event-Localized Multi-Turn Reinforcement Learning for Real-Time MOBA Commentary](https://mlsanity.com/paper/2609.38428/) (2026)
- [LIBERO-Agent: Evaluating General-Purpose Agents for Direct Embodied Manipulation](https://mlsanity.com/paper/2609.39507/) (2026)
- [The Evolution of Attention in Large Language Models: Mechanisms, Trade-offs, and Emerging Trends](https://mlsanity.com/paper/2609.39661/) (2026)
- [FIGS: Evaluating Multi-Turn Sycophancy Without Penalizing Empathy](https://mlsanity.com/paper/2609.39863/) (2026)
- [Efficient Expert-Parallel Communication on PCIe-Connected Consumer GPUs](https://mlsanity.com/paper/2609.40093/) (2026)
- [Forking: Sudden Overfitting Under Replay](https://mlsanity.com/paper/2610.00394/) (2026)
- [CompMat-Bench: Benchmarking AI Agents for Computational Materials Science](https://mlsanity.com/paper/2610.00636/) (2026)
- [AuraForge: Scaling Security Supervision for Training Coding Agents](https://mlsanity.com/paper/2610.00850/) (2026)

## Abstract (arXiv)

The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transfer bandwidth. Together, these compute, storage, and bandwidth demands constitute the primary bottleneck to further lowering deployment costs. To address this challenge, we introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. With its Causal Encoder-Decoder (CED) architecture, the model activates 16B parameters per token during decode but only 8B parameters during prefill, substantially improving cost efficiency for agentic workloads. To push the limits of KV cache compression, DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching. These designs reduce its global KV cache footprint (always in HBM) to 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash. Further, through a dedicated deployment optimization known as SWA Bounded Replay, DeepSeek-V4.1-Flash reduces its persistent KV cache footprint (always on SSD or in host memory) to roughly 1/8 of that of DeepSeek-V4-Flash. Despite its much smaller KV cache footprint, the model delivers substantially better performance than the baseline. In addition, we streamline the DeepSeek-V4 architecture and introduce several efficient architectural extensions. We pretrain DeepSeek-V4.1-Flash on a multimodal corpus comprising 45T tokens and conduct comprehensive post-training, yielding strong performance across diverse text-based and multimodal agentic scenarios. Model checkpoints are available at https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.