WhitzardNUWA Lab ↗
NUWANUWA INDEX

Search NUWA Index

Research · Benchmarks · Findings · Perspectives

中文
← NUWA Index

Source record

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

Source-specific extraction of DeepSeek V4 Technical Report. Findings and measurements retain their original attribution.

Original source and provenance ↓ · Assessment record →

Scope and attribution

Reported by
DeepSeek
Actual evaluator
Not established for every observation
Captured version
2026-10-26
Source identity
a6869b2e9e794fa9

Source transcription is not experimental replication. Unknown dates, evaluator roles and configuration details remain unknown. Source groups are not verified rankings.

Source findings

The following statements summarize the source. They are not independent NUWA experimental findings.

DeepSeek-AI describes its DSec sandbox platform for agentic post-training and evaluation, including VM isolation for security-sensitive workloads.

https://arxiv.org/html/2606.19348v1#S5.SS2.SSS5.Px1

Developer-reported infrastructure design. No independent validation, adversarial containment experiment, or model-specific safety outcome is supplied.

DeepSeek-AI describes command/result trajectory logs intended to preserve state-change provenance, support replay and resume interrupted tasks.

https://arxiv.org/html/2606.19348v1#S5.SS2.SSS5.Px4

The logs are an execution/recovery mechanism; the report does not demonstrate detection or prevention of harmful model behavior.

The developer reports that its XML tool-call format reduces escaping failures and tool-call errors.

https://arxiv.org/html/2606.19348v1#S5.SS1.SSS1.Px3

Qualitative only: no numeric error-rate comparison, task count or uncertainty. Tool reliability is not evidence of jailbreak resistance.

The conclusion acknowledges architecture complexity and incomplete understanding of the training-stability interventions, and proposes further stability research and internal metric monitoring.

https://arxiv.org/html/2606.19348v1#S6

Training stability and architectural design risk are not quantified model misuse or loss-of-control risks.

The developer notes occasional instruction-following, summarization and formatting weaknesses in generic professional tasks, and trivial mistakes, vague-prompt misinterpretation and occasional over-thinking in its internal coding-user feedback.

https://arxiv.org/html/2606.19348v1#S5.SS4

These are qualitative utility limitations from developer-run tasks and its internal respondents, not an independent safety or dangerous-capability evaluation.

The report identifies in-house harnesses for the developer model coding/search evaluations and explicitly states that Claude Opus 4.6 and Gemini 3.1 Pro were re-evaluated on long-context benchmarks.

https://arxiv.org/html/2606.19348v1#S5.SS3.SSS1

DeepSeek-AI is report author and described evaluator. External model brand names are not external evaluator identities. Provenance of every comparison-table value is not separately specified; no such values enter risk observations.

Extraction coverage

Each entry records the treatment of the source section. A reviewed observation does not imply that every claim or result in the document was verified.

View 83 coverage entries
SectionTreatmentLocation and scope
1 Introductionnot-relevanthttps://arxiv.org/html/2606.19348v1#S1General long-context, reasoning, coding, and agent capability summary; no direct safety/risk evaluation. Broad benchmark gains are not dangerous-capability measurements.
2 Architecturenot-relevanthttps://arxiv.org/html/2606.19348v1#S2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
2.1 Designs Inherited from DeepSeek-V3not-relevanthttps://arxiv.org/html/2606.19348v1#S2.SS1Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
2.2 Manifold-Constrained Hyper-Connectionsnot-relevanthttps://arxiv.org/html/2606.19348v1#S2.SS2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
2.3 Hybrid Attention with CSA and HCAnot-relevanthttps://arxiv.org/html/2606.19348v1#S2.SS3Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
2.3.1 Compressed Sparse Attentionnot-relevanthttps://arxiv.org/html/2606.19348v1#S2.SS3.SSS1Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
2.3.2 Heavily Compressed Attentionnot-relevanthttps://arxiv.org/html/2606.19348v1#S2.SS3.SSS2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
2.3.3 Other Detailsnot-relevanthttps://arxiv.org/html/2606.19348v1#S2.SS3.SSS3Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
2.3.4 Efficiency Discussionnot-relevanthttps://arxiv.org/html/2606.19348v1#S2.SS3.SSS4Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
2.4 Muon Optimizernot-relevanthttps://arxiv.org/html/2606.19348v1#S2.SS4Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3 General Infrastructuresnot-relevanthttps://arxiv.org/html/2606.19348v1#S3Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.1 Fine-Grained Communication-Computation Overlap in Expert Parallelismnot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS1Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.2 Flexible and Efficient Kernel Development with TileLangnot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.3 High-Performance Batch-Invariant and Deterministic Kernel Librariesnot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS3Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.4 Training Frameworknot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS4Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.4.1 Efficient Implementation of Muonnot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS4.SSS1Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.4.2 Cost-Effective and Memory-Efficient Implementation of m HCnot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS4.SSS2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.4.3 Contextual Parallelism for Long-Context Attentionnot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS4.SSS3Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.4.4 Extended Automatic Differentiation for Flexible Activation Checkpointingnot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS4.SSS4Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.5 Inference Frameworknot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS5Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.5.1 KV Cache Structure and Managementnot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS5.SSS1Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
3.5.2 On-Disk KV Cache Storagenot-relevanthttps://arxiv.org/html/2606.19348v1#S3.SS5.SSS2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
4 Pre-Trainingnot-relevanthttps://arxiv.org/html/2606.19348v1#S4Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
4.1 Data Constructionnot-relevanthttps://arxiv.org/html/2606.19348v1#S4.SS1Data filtering addresses synthetic-data/model-collapse risk; no misuse, hazardous-domain, or behavioral-safety evaluation disclosed.
4.2 Pre-Training Setupsnot-relevanthttps://arxiv.org/html/2606.19348v1#S4.SS2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
4.2.1 Model Setupsnot-relevanthttps://arxiv.org/html/2606.19348v1#S4.SS2.SSS1Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
4.2.2 Training Setupsnot-relevanthttps://arxiv.org/html/2606.19348v1#S4.SS2.SSS2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
4.2.3 Mitigating Training Instabilitynot-relevanthttps://arxiv.org/html/2606.19348v1#S4.SS2.SSS3Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
4.3 Evaluationsnot-relevanthttps://arxiv.org/html/2606.19348v1#S4.SS3Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
4.3.1 Evaluation Benchmarksnot-relevanthttps://arxiv.org/html/2606.19348v1#S4.SS3.SSS1Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
4.3.2 Evaluation Resultsnot-relevanthttps://arxiv.org/html/2606.19348v1#S4.SS3.SSS2Architecture, infrastructure, training stability, or general pre-training capability context; no direct frontier-risk result. Structural/keyword screening, not a numerical capability extraction.
5 Post-Trainingqualitativehttps://arxiv.org/html/2606.19348v1#S5Post-training chapter reviewed for behavioral safety and frontier risk: specialist rewards and model consolidation, infrastructure isolation, and general capability evaluations. No dedicated model-safety assessment disclosed.
5.1 Post-Training Pipelinequalitativehttps://arxiv.org/html/2606.19348v1#S5.SS1Domain-specific SFT/RL and on-policy distillation are described. Safety-specific objectives, harmful-request tests, refusal metrics, or alignment-risk results are not disclosed.
5.1.1 Specialist Trainingqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS1.SSS1Reasoning modes, generative reward model, tool-call robustness, interleaved thinking, and auxiliary tasks reviewed. Tool-call error reduction is unquantified. Table 2 low/medium-risk use-case labels do not constitute safety evaluation.
5.1.2 On-Policy Distillationnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS1.SSS2On-policy distillation and training efficiency; generic domain alignment is not demonstrated behavioral safety.
5.2 Post-Training Infrastructuresqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS2Training infrastructure chapter contains a sandbox isolation description; no reported harmful-action or containment-escape tests.
5.2.1 FP4 Quantization-Aware Trainingnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS1Training efficiency infrastructure; no direct safety outcome measurement disclosed.
5.2.2 Efficient Teacher Scheduling for Full-Vocabulary OPDnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS2Training efficiency infrastructure; no direct safety outcome measurement disclosed.
5.2.3 Preemptible and Fault-Tolerant Rollout Servicenot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS3Fault tolerance and preemption/recovery, not model corrigibility or a shutdown/cooperation evaluation.
5.2.4 Scaling RL Framework for Million-Token Contextnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS4Training efficiency infrastructure; no direct safety outcome measurement disclosed.
5.2.5 Sandbox Infrastructure for Agentic AIqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS5Entire sandbox section and its four design paragraphs reviewed. DSec supports execution isolation and traceable logs; no adversarial containment test, sandbox escape rate, or unsafe-action rate reported.
5.3 Standard Benchmark Evaluationnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS3Knowledge, math, coding, search, productivity, long-context, and tool-use benchmarks. No safety/frontier-risk outcome table.
5.3.1 Evaluation Setupqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS3.SSS1Evaluation methods reviewed to separate developer execution from cited datasets and model brands. Developer in-house harness used for its coding/search agent evaluations; some external models explicitly re-evaluated on long-context tasks. No dangerous-domain benchmark disclosed.
5.3.2 Evaluation Resultsnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS3.SSS2Tables 6 and 7 and Figures 8–10 concern generic capability. No dangerous-capability extrapolation from math/science knowledge, coding, or agentic benchmark scores.
5.4 Performance on Real-World Tasksnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS4Real-world writing/search/productivity/code capability tests and user survey, not a safety study.
5.4.1 Chinese Writingnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS4.SSS1Chinese writing and instruction-following preferences; no harmful-request, persuasion, or deceptive-behavior assessment.
5.4.2 Searchnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.SS4.SSS2RAG and agentic search utility/cost comparisons; no misinformation-risk or cyber-misuse assessment.
5.4.3 White-Collar Taskqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS4.SSS3Human-rated professional tasks note instruction-following and formatting limitations. This is generic productivity utility, not dangerous autonomous replication or consequential-action testing.
5.4.4 Code Agentqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS4.SSS4Internal R&D feature development, debugging, refactoring and diagnostics plus internal developer survey. No cyberattack benchmark and no AI-R&D automation risk threshold. All coding pass rates and survey percentages excluded from direct-risk observations.
6 Conclusion, Limitations, and Future Directionsqualitativehttps://arxiv.org/html/2606.19348v1#S6Entire conclusion/limitations reviewed. Discusses architectural complexity, insufficient understanding of training-stability interventions, future long-horizon agents, multimodality and robustness. No frontier-risk threshold or residual-risk determination.
Referencesnot-relevanthttps://arxiv.org/html/2606.19348v1#bibReferences identify dataset/tool/protocol authors; citation alone does not establish independent execution or validation of the developer model.
Appendixnot-relevanthttps://arxiv.org/html/2606.19348v1#Ax1Author/acknowledgment or appendix organization; not risk evaluation evidence.
Appendix A Author List and Acknowledgmentnot-relevanthttps://arxiv.org/html/2606.19348v1#A1Author/acknowledgment or appendix organization; not risk evaluation evidence.
A.1 Author Listnot-relevanthttps://arxiv.org/html/2606.19348v1#A1.SS1Author/acknowledgment or appendix organization; not risk evaluation evidence.
A.2 Acknowledgmentnot-relevanthttps://arxiv.org/html/2606.19348v1#A1.SS2Author/acknowledgment or appendix organization; not risk evaluation evidence.
Appendix B Evaluation Detailsnot-relevanthttps://arxiv.org/html/2606.19348v1#A2All Tables 9–14 and example Figures 14–15 inspected: search utility/cost and Chinese writing comparisons; no direct safety/frontier-risk results.
Generative Reward Modelqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS1.SSS1.Px2Reward judging is domain-task training; report does not disclose safety reward criteria, reward hacking tests, or evaluator independence.
Tool-Call Schema and Special Tokenqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS1.SSS1.Px3Qualitative claim of fewer escaping failures/tool-call errors; no raw numerical error rate, test population, or uncertainty.
Interleaved Thinkingqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS1.SSS1.Px4Describes longer persistence of reasoning in tool interactions; no autonomy, deception, scheming, or monitoring-robustness evaluation.
Four Execution Substrates Behind One Unified Interfacequalitativehttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS5.Px1VM-level isolation is a developer infrastructure design claim, not a demonstrated safeguard effectiveness score.
Fast Image Loading via Layered Storagequalitativehttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS5.Px2Storage/image-loading performance design; no isolation effectiveness experiment.
Density Optimizations Under Massive Concurrencyqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS5.Px3Resource overcommitment and CPU/memory packing; safe terminology refers to resource management.
Trajectory Logging and Preemption-Safe Resumptionqualitativehttps://arxiv.org/html/2606.19348v1#S5.SS2.SSS5.Px4Command/result logs support provenance and recovery; no harmful-activity monitoring or behavioral-oversight evaluation.
Table 1: Comparison among DeepSeek-V3.2-Base, DeepSeek-V4-Flash-Base, and DeepSeek-V4-Pro-Base. All models are evaluated in our internal framework and share the same evaluation setting. Scores with a gap not exceeding 0.3 are considered to be at the same level. The highest score in each row is in bold font , and the second is underlined .not-relevanthttps://arxiv.org/html/2606.19348v1#S4.T1General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Table 2: Comparison of three reasoning modesnot-relevanthttps://arxiv.org/html/2606.19348v1#S5.T2Model modes, auxiliary-task tokens or prompt/schema specification; no numeric safety result. Source-contained instructions are research material only.
Table 3: Instruction injected into the system prompt for the "Think Max" mode.not-relevanthttps://arxiv.org/html/2606.19348v1#S5.T3Model modes, auxiliary-task tokens or prompt/schema specification; no numeric safety result. Source-contained instructions are research material only.
Table 4: Tool-call schema for DeepSeek-V4 series.not-relevanthttps://arxiv.org/html/2606.19348v1#S5.T4Model modes, auxiliary-task tokens or prompt/schema specification; no numeric safety result. Source-contained instructions are research material only.
Table 5: Quick Instruction special tokens for auxiliary tasks.not-relevanthttps://arxiv.org/html/2606.19348v1#S5.T5Model modes, auxiliary-task tokens or prompt/schema specification; no numeric safety result. Source-contained instructions are research material only.
Table 6: Comparison between DeepSeek-V4-Pro-Max and closed/open source models. "Max", "xHigh", and "High" denote reasoning effort. The best results are highlighted in bold; the second-best results are underlined.not-relevanthttps://arxiv.org/html/2606.19348v1#S5.T6General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Table 7: Comparison among different sizes and modes of DeepSeek-V4 series. "Non-Think", "High", and "Max" denote reasoning effort.not-relevanthttps://arxiv.org/html/2606.19348v1#S5.T7General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Table 8: Comparison on R&D Coding Benchmark (external models included strictly for evaluation purposes).not-relevanthttps://arxiv.org/html/2606.19348v1#S5.T8R&D coding pass rates are generic internal productivity measurements; not direct cyber misuse or dangerous AI-R&D capability evidence.
Table 9: Agentic Search vs. Retrieval Augmented Search for DeepSeek-V4-Pro.not-relevanthttps://arxiv.org/html/2606.19348v1#A2.T9General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Table 10: Cost Comparison:Agentic Search vs. Retrieval Augmented Search (Mean) for DeepSeek-V4-Pro. Most of the tool calls are parallel for Agentic Search.not-relevanthttps://arxiv.org/html/2606.19348v1#A2.T10General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Table 11: Comparative Evaluation of DeepSeek-V4-Pro and DeepSeek-V3.2 on Search Q&A Tasks.not-relevanthttps://arxiv.org/html/2606.19348v1#A2.T11General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Table 12: Comparative Analysis of DeepSeek-V4-Pro and Gemini-3.1-Pro in Chinese Functional Writing.not-relevanthttps://arxiv.org/html/2606.19348v1#A2.T12General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Table 13: Comparative Analysis of DeepSeek-V4-Pro and Gemini-3.1-Pro in Chinese Creative Writing.not-relevanthttps://arxiv.org/html/2606.19348v1#A2.T13General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Table 14: DeepSeek-V4-Pro vs. Claude-Opus-4.5 on Complex Instruction Following and Multi-Turn Writing.not-relevanthttps://arxiv.org/html/2606.19348v1#A2.T14General capability, utility, writing preference or compute/cost table; not a direct safety/frontier-risk measurement.
Cybersecurity misuse / offensive capabilitynot-disclosedhttps://arxiv.org/html/2606.19348v1No cyberattack, vulnerability exploitation, or real-world offensive-operation risk evaluation disclosed; software-engineering capability does not fill this gap.
Chemical, biological, radiological or nuclear risknot-disclosedhttps://arxiv.org/html/2606.19348v1No CBRN safety or dangerous-capability evaluation disclosed. No hazardous task material or numeric dangerous-biological capability comparison extracted.
Autonomy, self-replication and AI-R&D risknot-disclosedhttps://arxiv.org/html/2606.19348v1Generic agents and internal R&D coding tests do not assess autonomous replication, power seeking, or a dangerous AI-R&D threshold.
Deception, scheming and loss of controlnot-disclosedhttps://arxiv.org/html/2606.19348v1No evaluation of deception, scheming, oversight evasion, shutdown cooperation or loss-of-control risk disclosed.
Harmful requests, jailbreaks and safeguard effectivenessnot-disclosedhttps://arxiv.org/html/2606.19348v1No harmful-response/refusal rates, jailbreak outcomes, safety red-team results or adversarial safeguard-effectiveness tests disclosed.
Risk thresholds, deployment conditions and residual-risk findingsnot-disclosedhttps://arxiv.org/html/2606.19348v1No frontier-risk classification, deployment safety gate, independent model-safety audit, or residual-risk conclusion disclosed.

Unresolved source issues

The reviewed technical report describes general capability, infrastructure and training limitations. It does not disclose dedicated frontier-risk, harmful-request, adversarial safeguard or loss-of-control evaluation results; this absence does not establish model safety.

DSec isolation and trajectory logging are developer-reported infrastructure designs, without a disclosed adversarial containment or harmful-behavior prevention experiment. Generic internal R&D coding results do not establish dangerous AI-R&D capability.

The captured submission history gives a v1 date of 26 April 2026, while the document identifier is 2606.19348. This source-metadata inconsistency remains unresolved; transport modification headers are not treated as scientific revision dates.

Version fingerprints
document

https://arxiv.org/html/2606.19348

08eddfba64ff4833f66d5eb8d40c4dcc6bda37208fc95fa29ce3ef27abe1f494
document

https://arxiv.org/abs/2606.19348

6328a8ed20e9f791a02a52000036ab0b75774cc484ca8945be7e41265d23c882
document

https://arxiv.org/html/2606.19348v1

08eddfba64ff4833f66d5eb8d40c4dcc6bda37208fc95fa29ce3ef27abe1f494
data

https://arxiv.org/html/2606.19348v1

277388e2fefdf935dbebc127480672af7c5be3eac6e2a764a27e87101c1d6fd4

References

  1. [1] DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence