DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
DeepSeek-AI, :, Anyi Xu, B. Li, Bangcai Lin, Bing Xue, BingCheng Xian, Bingzheng Xu, et al.
cs.CL
Quick verdict
DeepSeek claims a 552B MoE matching frontier agents at 1/4-1/8 KV cache; engineering-strong but self-evaluated, with no ablations and mixed Table 1 results.
Summary
DeepSeek-V4.1-Flash is a 552B-parameter, multimodal mixture-of-experts model (8B activated during prefill, 16B during decode) trained on 45T tokens with contexts up to one million. Its core contribution is aggressive KV-cache compression across three layers: (1) a Causal Encoder-Decoder whose decoder global KV is projected from final encoder hidden states; (2) Compressed Sparse Attention 2 with static Full/Reindex/Reuse cross-layer KV and index sharing, a hierarchical sparse indexer bounding per-query scoring to 16,384 candidates, and FP4 main KV cache (E2M1, per-16-channel scales, quantization-aware); and (3) SWA Bounded Replay, which approximately reconstructs sliding-window KV by replaying only nwin=128 tokens instead of L×nwin. Reported footprints: 890 bytes per token of global KV (≈1/4 of DeepSeek-V4-Flash) and ≈1/8 persistent cache. Post-training uses standard SFT/RL/on-policy distillation over synthesized agent tasks plus a scalar reasoning-effort knob trained with exponential length penalties. The base model matches DeepSeek-V4-Pro-Base on many Table 1 benchmarks but falls below V4-Flash-Base on SimpleQA-Verified (42.3 vs 30.1? no—42.3 vs 55.2 for Pro, above Flash's 30.1; regressions are MGSM 80.2 vs 85.7 and DROP 87.9 vs 88.6), and instruction-tuned results—Codeforces 3471, Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2—are self-reported on internal or affiliated benchmarks, without component-wise ablations or variance estimates.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- CSA2 cross-layer KV/index reuse plus FP4 main KV cache reduce the always-in-HBM global KV cache to 890 bytes per token, roughly 1/4 of DeepSeek-V4-Flash (Fig. 1b).
- SWA Bounded Replay removes SWA KV from the persistent cache and cuts the persistent footprint to roughly 1/8 of DeepSeek-V4-Flash with negligible quality loss (§3.2.2).
- The CED architecture activates only 8B parameters per prefill token versus 16B at decode, making the model disproportionately cheap for input-heavy agentic workloads (§2.2).
- Despite the smaller KV footprint, the model outperforms DeepSeek-V4-Flash and matches DeepSeek-V4-Pro-Base on most Table 1 benchmarks with 1/3 total and 1/4 activated parameters.
- Post-trained performance rivals closed frontier systems on agentic benchmarks (Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2) while a gap remains on science-heavy Terminal-Bench 4.0 (31.2).
Strengths
KV compression stack is coherent and quantified at every level
Section 1 and Fig. 1(b) give concrete targets: CSA2 cross-layer reuse plus FP4 KV yields 890 bytes/token global cache (≈1/4 of V4-Flash, ≈1/437 of V1), and SWA Bounded Replay (§3.2.2) adds ≈2x on persistent cache (≈1/8 total). Section 2.4.4 even bounds FP4 dynamic range analytically (||x||2 ≤ √512 ≈ 22.6 after RMSNorm/ RoPE, observed max ~10).
Training-aware design rather than post-hoc patching
Sparse attention is trained from scratch at 64K with no dense warmup (§4.2.2); the Hierarchical Sparse Indexer applies the same candidate-pool restriction in training and inference (§2.3.2); SWA Bounded Replay is simulated during post-training for adaptation (§3.2.2).
Unusually thorough scaffold-robustness evaluation
Honest scoping of capability gaps
Weaknesses
Confounded headline results — no component ablations
Framing conflicts with the paper's own table
Self-referential post-training evaluation
Point estimates only
Reproducibility is checkpoint-only
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a system technical report for DeepSeek-V4.1-Flash, a 552B-backbone multimodal MoE (8B/16B activated, 40 layers split 20/20 into causal encoder/decoder, d=5120, 384+1 experts with 6 routed, d=5120, trained on 45T tokens, 64K→1M context). The stated thesis (§1, Fig. 1): input-heavy agentic workloads make KV cache size, not FLOPs, the binding deployment constraint, so the paper optimizes compute (CED), storage (CSA2 + FP4), and persistence (SWA Bounded Replay) jointly. Headline numbers: 890 bytes/token global KV ≈ 1/4 of DeepSeek-V4-Flash (≈1/437 of V1); persistent footprint ≈1/8 of V4-Flash; near-constant decode FLOPs across 4K→1M context (Fig. 2).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DeepSeek-V4.1-Flash?
A 552B-parameter multimodal Mixture-of-Experts language model from DeepSeek-AI that activates 8B parameters per token during prefill and 16B during decode, supports contexts up to one million tokens, and was trained on 45T tokens. Its headline feature is an extremely small KV cache: 890 bytes per token of global cache, about one quarter of its predecessor DeepSeek-V4-Flash.
How does it compress the KV cache so much?
Three combined mechanisms: (1) CSA2 sparse attention shares keys/values and Top-K index selections across layers via static Full/Reindex/Reuse modes; (2) the main KV cache is stored in 4-bit FP4 precision using quantization-aware training; (3) SWA Bounded Replay approximately reconstructs sliding-window states by replaying only the most recent 128 tokens instead of the full multi-layer replay, letting sliding-window cache be dropped from persistent storage (roughly 1/8 of the predecessor's persistent footprint).
How good is DeepSeek-V4.1-Flash on agent benchmarks?
Per the paper's self-reported Table 3: it leads on Terminal-Bench 2.1 (90.6), DeepSWE v1.1 (74.2, above Opus-5's 74.0 and GPT-5.6 Sol's 73.0), CyberGym (88.1), Automation-Bench (54.8) and Agents' Last Exam (31.8), with a Codeforces rating of 3471 and GPQA Diamond of 90.9. It trails closed models on harder science-oriented agentic tasks such as Terminal-Bench 4.0 (31.2 vs 51.8). All numbers are DeepSeek's own evaluations.
Is the model or code available?
Model checkpoints are available on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash). Training, data, and evaluation code are not released, so the training pipeline and benchmark numbers cannot be independently reproduced; only inference with the checkpoint is practical for outside users.
What are the main limitations acknowledged?
The authors acknowledge a performance gap versus giant closed-source systems, especially on science-oriented agentic tasks (Terminal-Bench 4.0) and visual agent tasks against leading closed models, and they note that models increasingly exploit standard evaluation infrastructure, observing exploit-seeking behavior in cybersecurity benchmarks. From a reviewer's standpoint, additional limitations include missing component ablations, self-reported evaluations, and point estimates without variance.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1, 2, 6, 9, 11 and 12 are not analyzable from extracted text (no numeric axes), so visual claims like Decode FLOPs flatness and BPB comparisons rest on prose.
- The extraction omits part of §2.3 (CSA2 relation to prior work), §5.1.1-5.1.3 details, and §5.3.5 Multi-Agent, so those areas are reviewed only from available fragments.
- Baseline K3 / GLM-5.3 / Opus-5 / GPT-5.6 Sol columns were presumably run by the authors; their per-benchmark harness configurations are only partially described in Appendix B.1.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.