Papersby GlyphDeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression research paper by DeepSeek, 2026
DeepSeek · Inference and efficiency · Sep 17, 2026 · 158 citations · 188 upvotes
What it shows
A 552B mixture-of-experts model with a 1M-token context, built to shrink the KV cache for agent workloads.
By DeepSeek-AI, Anyi Xu, B. Li and 589 more · arXiv 2609.19969 · PDF
Summarised by hand from the abstract.
The model it introduced
On models.fru.devMore from DeepSeek
All 22DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive GenerationDSpark enhances LLM inference speed by combining parallel draft generation with adaptive verification that reduces waste and improves throughput in high-concurrency settings.
DeepSeek77 citations49 upvotesRead →DualPath: Breaking the Storage Bandwidth Bottleneck in Agentic LLM InferenceDualPath addresses KV-cache storage I/O bottlenecks in multi-turn LLM inference by introducing dual-path loading and dynamic load balancing across prefill and decode engines.DeepSeek22 citations55 upvotesRead →DeepSeek-OCR 2: Visual Causal FlowDeepSeek-OCR 2 introduces DeepEncoder V2 that dynamically reorders visual tokens based on semantic content, enabling more human-like causal reasoning in 2D image understanding through cascaded 1D causal structures.DeepSeek79 citations73 upvotesRead →mHC: Manifold-Constrained Hyper-ConnectionsManifold-Constrained Hyper-Connections (mHC) stabilize and scale residual connection architectures by restoring identity mapping properties through manifold projection and infrastructure optimization.DeepSeek108 citations334 upvotesRead →DeepSeek-V3.2: Pushing the Frontier of Open Large Language ModelsDeepSeek-V3.2 introduces DeepSeek Sparse Attention and a scalable reinforcement learning framework, achieving superior reasoning and performance compared to GPT-5 and Gemini-3.0-Pro in complex reasoning tasks.DeepSeek887 citations275 upvotesRead →DeepSeekMath-V2: Towards Self-Verifiable Mathematical ReasoningA self-verifying large language model for theorem proving improves mathematical reasoning by incentivizing rigorous step-by-step derivations and achieves high scores in international competitions.DeepSeek83 citations98 upvotesRead →DeepSeek-OCR: Contexts Optical CompressionDeepSeek-OCR uses optical 2D mapping to compress long contexts, achieving high OCR precision with reduced vision tokens and demonstrating practical value in document processing.DeepSeek200 citations98 upvotesRead →Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI ArchitecturesDeepSeek-V3 addresses hardware limitations through MLA, MoE, FP8 training, and Multi-Plane Network Topology, enabling efficient large-scale LLM training and inference.DeepSeek116 citations77 upvotesRead →
Releases that name it
From releases.fru.devDeepSeek V4.1 Flash 1017 Ultrafast now available on AI Gatewayvercelreleases.fru.dev →DeepSeek V4.1 Flash (1017) support in AIP through Fireworkspalantirreleases.fru.dev →New model: DeepSeek V4.1 Flashmodularreleases.fru.dev →DeepSeek V4.1 Flash is now available in Token Passvercelreleases.fru.dev →DeepSeek V4.1 Flash (1017) is now available on Unity Gatewaydatabricksreleases.fru.dev →Workers AI - DeepSeek V4.1 Flash 1017 now available on Workers AIcloudflarereleases.fru.dev →DeepSeek V4.1 Flash is scheduled for retirementdatabricksreleases.fru.dev →DeepSeek V4.1 Flash 1017 now available on AI Gatewayvercelreleases.fru.dev →DeepSeek-V4.1-Flash is available in Public Previewgoogle cloudreleases.fru.dev →DeepSeek V4.1 Flash supports priority pay-per-token and reserved provisioned throughputdatabricksreleases.fru.dev →Training New training model: DeepSeek V4.1 Flash DeepSeek V4.1 Flash is now available for LoRA training on...fireworks aireleases.fru.dev →Import DeepSeek V4.1 Flash into OCI Generative AIoraclereleases.fru.dev →
Other inference and efficiency papers
TopicEfficient Memory Management for Large Language Model Serving with PagedAttentionvLLM manages the KV cache like pages of virtual memory, serving models with 2 to 4 times the throughput.
UC Berkeley9,965 citations72 upvotesRead →FlashAttention: Fast and Memory-Efficient Exact Attention with IO-AwarenessFlashAttention computes exact attention with far fewer GPU memory reads and writes, making long sequences faster.
Stanford University5,830 citations16 upvotesRead →Group Sequence Policy OptimizationGroup Sequence Policy Optimization (GSPO) is a reinforcement learning algorithm that improves training efficiency and performance of large language models by using sequence-level importance ratios and operations.
Alibaba (Qwen)814 citations323 upvotesRead →SmolVLM: Redefining small and efficient multimodal modelsSmolVLM, a series of compact multimodal models, achieves high performance with minimal GPU memory usage, making efficient deployment on mobile and edge devices possible.
Hugging Face330 citations212 upvotesRead →Chronos-2: From Univariate to Universal ForecastingChronos-2, a pretrained model with a group attention mechanism, achieves state-of-the-art performance in zero-shot univariate, multivariate, and covariate-informed forecasting tasks.
Amazon276 citations28 upvotesRead →ACON: Optimizing Context Compression for Long-horizon LLM AgentsAgent Context Optimization (ACON) compresses context in large language models for efficient long-horizon tasks by analyzing failure cases and distilling the compressor into smaller models.
Microsoft162 citations36 upvotesRead →
Amazon276 citations28 upvotesRead →ACON: Optimizing Context Compression for Long-horizon LLM AgentsAgent Context Optimization (ACON) compresses context in large language models for efficient long-horizon tasks by analyzing failure cases and distilling the compressor into smaller models.About this paper
- Citations
- 158, 6 influential · Semantic Scholar
- Upvotes
- 188 · Hugging Face
- Lab
- DeepSeek · on Companies · on breaches · on Rounds · on labs · on Leaders · on Arena · on legislations · on Paydays · on Metered · on problems · on repos · on socials · on standards · on terms
Changes
Influential citations rose from 3 to 6.
Citations rose from 60 to 158.
Influential citations rose from 0 to 3.
Citations rose from 36 to 60.
Citations rose from 7 to 36.
Hugging Face upvotes rose from 178 to 188.
Hugging Face upvotes rose from 175 to 178.
Added to the list.