"DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence" is the technical report published by DeepSeek for the preview release of its DeepSeek-V4 series. arXiv lists it as submitted on April 26, 2026 (arXiv:2606.19348), and the model cards on Hugging Face link it as the series' technical report (Source: arxiv.org; huggingface.co). It covers two mixture-of-experts language models, DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated), both supporting a context length of one million tokens, and their base checkpoints.
Architecture and training
The report presents three changes relative to DeepSeek-V3: a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA); Manifold-Constrained Hyper-Connections (mHC), which replace conventional residual connections; and the Muon optimizer. It retains the DeepSeekMoE design and multi-token prediction from V3. DeepSeek states that at a one-million-token context V4-Pro requires 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2.
Both models were pre-trained on more than 32T tokens. The report describes the corpus as built on the V3 pre-training data with added filtering of auto-generated web content, a larger multilingual corpus, long-document curation favouring scientific papers and technical reports, and agentic data added in mid-training. Two techniques, Anticipatory Routing and SwiGLU Clamping, are described as mitigations for training instability; the conclusion says their underlying principles "remain insufficiently understood."
Post-training follows the DeepSeek-V3.2 pipeline with one substitution: the mixed reinforcement-learning stage is replaced by on-policy distillation. Domain specialists are first trained by supervised fine-tuning and reinforcement learning with GRPO, then consolidated into one model through distillation. The released models expose three reasoning modes, Non-think, High and Max; the report calls the Max mode of each model DeepSeek-V4-Pro-Max and DeepSeek-V4-Flash-Max. Infrastructure sections cover expert-parallel communication overlap, TileLang kernels, deterministic kernel libraries, FP4 quantization-aware training, on-disk KV cache storage and a sandbox service for agentic training.
Evaluation approach
Base models are compared with DeepSeek-V3.2-Base in an internal framework under one setting (Table 1), across world knowledge, language understanding and reasoning, code and mathematics, and long context.
Post-trained models are evaluated on knowledge and reasoning benchmarks (MMLU-Pro, GPQA Diamond, Humanity's Last Exam, SimpleQA-Verified, Chinese-SimpleQA, LiveCodeBench-v6, an internal Codeforces benchmark, HMMT 2026 February, IMOAnswerBench, Apex and Apex Shortlist), one-million-token context tasks (OpenAI MRCR and CorpusQA), and agentic tasks (Terminal Bench 2.0, SWE Verified, SWE Pro, SWE Multilingual, BrowseComp, Humanity's Last Exam with tools, GDPval-AA, MCPAtlas Public and Toolathlon). Reasoning and knowledge tasks use temperature 1.0 with context windows of 8K, 128K and 384K tokens for the Non-think, High and Max modes. Code-agent tasks use an internal framework with a bash tool and a file-edit tool, up to 500 steps and a 512K-token context; search-agent tasks use an in-house harness with web search and Python. The Codeforces rating is computed from 14 Division 1 contests (114 problems) held between May and November 2025. Table 6 compares DeepSeek-V4-Pro-Max with Claude Opus 4.6, GPT-5.4, Gemini 3.1 Pro, Kimi K2.6 and GLM-5.1; Table 7 compares the three reasoning modes of both V4 models. The report says some Kimi K2.6 and GLM-5.1 cells are blank because those APIs were too busy to respond, that GPT-5.4 was not run on the long-context tasks because its API failed on a large share of queries, and that Claude Opus 4.6 and Gemini 3.1 Pro were re-evaluated by DeepSeek on the long-context tasks.
A further section reports internal evaluations on the main uses of DeepSeek's API and chatbot: pairwise Chinese functional and creative writing comparisons with Gemini 3.1 Pro, a comparison with Claude Opus 4.5 on complex instruction following and multi-turn writing, retrieval-augmented and agentic search, a human-judged set of 30 Chinese white-collar tasks against Claude Opus 4.6, and a 30-task coding benchmark drawn from internal R&D work. It also reports formal-mathematics results on Putnam problems in Lean, and a survey of 85 DeepSeek developers and researchers on whether V4-Pro was ready to be their default coding model.
Findings as stated by DeepSeek
DeepSeek describes V4-Pro-Max as the strongest open model on knowledge benchmarks while trailing Gemini 3.1 Pro, and as matching closed models on several reasoning benchmarks. On code-agent tasks the report says V4-Pro is comparable to Kimi K2.6 and GLM-5.1 and that "all these open models still lag behind their closed-source counterparts." On the one-million-token MRCR task it places V4-Pro ahead of Gemini 3.1 Pro and behind Claude Opus 4.6, with retrieval stable to 128K tokens and degrading beyond. On the internal writing comparison restricted to the hardest prompts, the report states that Claude Opus 4.5 retains an advantage over V4-Pro. The report notes environment-related issues in Terminal Bench 2.0 raised by the GLM-5.1 developers and reports the original dataset for consistency.
The conclusion lists limitations: an architecture the authors call "relatively complex" because preliminarily validated components were retained to reduce risk, training-stability techniques whose mechanisms are not understood, and the absence of multimodal capability, which DeepSeek says it is working on.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| AGIEval | DeepSeek-V3.2-Base | 80.1 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| AGIEval | DeepSeek-V4-Flash-Base | 82.6 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| AGIEval | DeepSeek-V4-Pro-Base | 83.1 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| MMLU-Redux | DeepSeek-V3.2-Base | 87.5 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| MMLU-Redux | DeepSeek-V4-Flash-Base | 89.4 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| MMLU-Redux | DeepSeek-V4-Pro-Base | 90.8 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| C-Eval | DeepSeek-V3.2-Base | 90.4 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| C-Eval | DeepSeek-V4-Flash-Base | 92.1 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| C-Eval | DeepSeek-V4-Pro-Base | 93.1 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| CMMLU | DeepSeek-V3.2-Base | 88.9 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| CMMLU | DeepSeek-V4-Flash-Base | 90.4 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| CMMLU | DeepSeek-V4-Pro-Base | 90.8 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| MultiLoKo | DeepSeek-V3.2-Base | 38.7 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| MultiLoKo | DeepSeek-V4-Flash-Base | 42.2 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| MultiLoKo | DeepSeek-V4-Pro-Base | 51.1 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| SuperGPQA | DeepSeek-V3.2-Base | 45.0 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| SuperGPQA | DeepSeek-V4-Flash-Base | 46.5 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| SuperGPQA | DeepSeek-V4-Pro-Base | 53.9 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| FACTS Parametric | DeepSeek-V3.2-Base | 27.1 | base model; 25-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| FACTS Parametric | DeepSeek-V4-Flash-Base | 33.9 | base model; 25-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| FACTS Parametric | DeepSeek-V4-Pro-Base | 62.6 | base model; 25-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| TriviaQA | DeepSeek-V3.2-Base | 83.3 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| TriviaQA | DeepSeek-V4-Flash-Base | 82.8 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| TriviaQA | DeepSeek-V4-Pro-Base | 85.6 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| BBH | DeepSeek-V3.2-Base | 87.6 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| BBH | DeepSeek-V4-Flash-Base | 86.9 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| BBH | DeepSeek-V4-Pro-Base | 87.5 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| DROP | DeepSeek-V3.2-Base | 88.2 | base model; 1-shot; F1; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| DROP | DeepSeek-V4-Flash-Base | 88.6 | base model; 1-shot; F1; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| DROP | DeepSeek-V4-Pro-Base | 88.7 | base model; 1-shot; F1; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| HellaSwag | DeepSeek-V3.2-Base | 86.4 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| HellaSwag | DeepSeek-V4-Flash-Base | 85.7 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| HellaSwag | DeepSeek-V4-Pro-Base | 88.0 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| WinoGrande | DeepSeek-V3.2-Base | 78.9 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| WinoGrande | DeepSeek-V4-Flash-Base | 79.5 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| WinoGrande | DeepSeek-V4-Pro-Base | 81.5 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| CLUEWSC | DeepSeek-V3.2-Base | 83.5 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| CLUEWSC | DeepSeek-V4-Flash-Base | 82.2 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| CLUEWSC | DeepSeek-V4-Pro-Base | 85.2 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| BigCodeBench | DeepSeek-V3.2-Base | 63.9 | base model; 3-shot; Pass@1; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| BigCodeBench | DeepSeek-V4-Flash-Base | 56.8 | base model; 3-shot; Pass@1; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| BigCodeBench | DeepSeek-V4-Pro-Base | 59.2 | base model; 3-shot; Pass@1; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| MGSM | DeepSeek-V3.2-Base | 81.3 | base model; 8-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| MGSM | DeepSeek-V4-Flash-Base | 85.7 | base model; 8-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| MGSM | DeepSeek-V4-Pro-Base | 84.4 | base model; 8-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| CMath | DeepSeek-V3.2-Base | 92.6 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| CMath | DeepSeek-V4-Flash-Base | 93.6 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| CMath | DeepSeek-V4-Pro-Base | 90.9 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| LongBench-V2 | DeepSeek-V3.2-Base | 40.2 | base model; 1-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4 technical report (DeepSeek's own earlier base model, same internal framework) | 4.3 Table 1 |
| LongBench-V2 | DeepSeek-V4-Flash-Base | 44.7 | base model; 1-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| LongBench-V2 | DeepSeek-V4-Pro-Base | 51.5 | base model; 1-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3 Table 1 |
| Chinese-SimpleQA | Claude Opus 4.6 | 76.4 | Pass@1; reasoning effort 'Max'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Chinese-SimpleQA | GPT-5.4 | 76.8 | Pass@1; reasoning effort 'xHigh'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Chinese-SimpleQA | Gemini 3.1 Pro | 85.9 | Pass@1; reasoning effort 'High'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Chinese-SimpleQA | Kimi K2.6 Thinking | 75.9 | Pass@1; reasoning effort 'Thinking'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Chinese-SimpleQA | GLM-5.1 Thinking | 75.0 | Pass@1; reasoning effort 'Thinking'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| IMOAnswerBench | Claude Opus 4.6 | 75.3 | Pass@1; reasoning effort 'Max'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| IMOAnswerBench | GPT-5.4 | 91.4 | Pass@1; reasoning effort 'xHigh'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| IMOAnswerBench | Gemini 3.1 Pro | 81.0 | Pass@1; reasoning effort 'High'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| IMOAnswerBench | Kimi K2.6 Thinking | 86.0 | Pass@1; reasoning effort 'Thinking'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| IMOAnswerBench | GLM-5.1 Thinking | 83.8 | Pass@1; reasoning effort 'Thinking'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| CorpusQA 1M | Claude Opus 4.6 | 71.7 | ACC; reasoning effort 'Max'; re-evaluated by DeepSeek to standardize configuration; reported by DeepSeek in the DeepSeek-V4 technical report | 5.3.2 Table 6 |
| CorpusQA 1M | Gemini 3.1 Pro | 53.8 | ACC; reasoning effort 'High'; re-evaluated by DeepSeek to standardize configuration; reported by DeepSeek in the DeepSeek-V4 technical report | 5.3.2 Table 6 |
| Toolathlon | Claude Opus 4.6 | 47.2 | Pass@1; reasoning effort 'Max'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Toolathlon | GPT-5.4 | 54.6 | Pass@1; reasoning effort 'xHigh'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Toolathlon | Gemini 3.1 Pro | 48.8 | Pass@1; reasoning effort 'High'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Toolathlon | Kimi K2.6 Thinking | 50.0 | Pass@1; reasoning effort 'Thinking'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Toolathlon | GLM-5.1 Thinking | 40.7 | Pass@1; reasoning effort 'Thinking'; reported by DeepSeek in the DeepSeek-V4 technical report; source of comparison figure not stated | 5.3.2 Table 6 |
| Chinese-SimpleQA | DeepSeek-V4-Flash (Non-Think) | 71.5 | Pass@1; Non-think mode (8K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| Chinese-SimpleQA | DeepSeek-V4-Flash (High) | 73.2 | Pass@1; High reasoning effort (128K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| Chinese-SimpleQA | DeepSeek-V4-Flash-Max | 78.9 | Pass@1; Max reasoning effort (384K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| Chinese-SimpleQA | DeepSeek-V4-Pro (Non-Think) | 75.8 | Pass@1; Non-think mode (8K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| Chinese-SimpleQA | DeepSeek-V4-Pro (High) | 77.7 | Pass@1; High reasoning effort (128K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| Chinese-SimpleQA | DeepSeek-V4-Pro-Max | 84.4 | Pass@1; Max reasoning effort (384K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 and Table 6 |
| IMOAnswerBench | DeepSeek-V4-Flash (Non-Think) | 41.9 | Pass@1; Non-think mode (8K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| IMOAnswerBench | DeepSeek-V4-Flash (High) | 85.1 | Pass@1; High reasoning effort (128K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| IMOAnswerBench | DeepSeek-V4-Flash-Max | 88.4 | Pass@1; Max reasoning effort (384K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| IMOAnswerBench | DeepSeek-V4-Pro (Non-Think) | 35.3 | Pass@1; Non-think mode (8K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| IMOAnswerBench | DeepSeek-V4-Pro (High) | 88.0 | Pass@1; High reasoning effort (128K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 |
| IMOAnswerBench | DeepSeek-V4-Pro-Max | 89.8 | Pass@1; Max reasoning effort (384K context window), temperature 1.0; developer-reported | 5.3.2 Table 7 and Table 6 |
| CorpusQA 1M | DeepSeek-V4-Flash (Non-Think) | 15.5 | ACC; Non-think mode (8K context window), temperature 1.0; 1M-token context; developer-reported | 5.3.2 Table 7 |
| CorpusQA 1M | DeepSeek-V4-Flash (High) | 59.3 | ACC; High reasoning effort (128K context window), temperature 1.0; 1M-token context; developer-reported | 5.3.2 Table 7 |
| CorpusQA 1M | DeepSeek-V4-Flash-Max | 60.5 | ACC; Max reasoning effort (384K context window), temperature 1.0; 1M-token context; developer-reported | 5.3.2 Table 7 |
| CorpusQA 1M | DeepSeek-V4-Pro (Non-Think) | 35.6 | ACC; Non-think mode (8K context window), temperature 1.0; 1M-token context; developer-reported | 5.3.2 Table 7 |
| CorpusQA 1M | DeepSeek-V4-Pro (High) | 56.5 | ACC; High reasoning effort (128K context window), temperature 1.0; 1M-token context; developer-reported | 5.3.2 Table 7 |
| CorpusQA 1M | DeepSeek-V4-Pro-Max | 62.0 | ACC; Max reasoning effort (384K context window), temperature 1.0; 1M-token context; developer-reported | 5.3.2 Table 7 and Table 6 |
| Toolathlon | DeepSeek-V4-Flash (Non-Think) | 40.7 | Pass@1; Non-Think mode; developer-reported | 5.3.2 Table 7 |
| Toolathlon | DeepSeek-V4-Flash (High) | 43.5 | Pass@1; High reasoning effort; developer-reported | 5.3.2 Table 7 |
| Toolathlon | DeepSeek-V4-Flash-Max | 47.8 | Pass@1; Max reasoning effort; developer-reported | 5.3.2 Table 7 |
| Toolathlon | DeepSeek-V4-Pro (Non-Think) | 46.3 | Pass@1; Non-Think mode; developer-reported | 5.3.2 Table 7 |
| Toolathlon | DeepSeek-V4-Pro (High) | 49.0 | Pass@1; High reasoning effort; developer-reported | 5.3.2 Table 7 |
| Toolathlon | DeepSeek-V4-Pro-Max | 51.8 | Pass@1; Max reasoning effort; developer-reported | 5.3.2 Table 7 and Table 6 |
| Terminal-Bench 2.0 Verified subset | DeepSeek-V4-Pro | approximately 72.0 | Terminal-Bench 2.0 Verified subset; internal framework with bash and file-edit tools; reasoning mode not stated; developer-reported | 5.3.1 |
| Putnam-2025 | DeepSeek-V4 | 120/120 | hybrid formal-informal pipeline with substantial compute scaling; Lean v4.28.0-rc1; verified by Comparator; model variant not stated; developer-reported | 5.3.2 Figure 8 caption |
| Codeforces human leaderboard rank | DeepSeek-V4-Pro-Max | 23rd | rank among human candidates on the Codeforces leaderboard implied by the 3206 rating; developer-reported | 5.3.2 Reasoning |
| R&D Coding Benchmark (internal) | Claude Haiku 4.5 | 13% | pass rate on 30 retained tasks from internal DeepSeek R&D workloads, human-annotated rubrics; reported by DeepSeek in the DeepSeek-V4 technical report | 5.4.4 Table 8 |
| R&D Coding Benchmark (internal) | Claude Sonnet 4.5 | 47% | pass rate on 30 retained tasks from internal DeepSeek R&D workloads, human-annotated rubrics; reported by DeepSeek in the DeepSeek-V4 technical report | 5.4.4 Table 8 |
| R&D Coding Benchmark (internal) | DeepSeek-V4-Pro-Max | 67% | pass rate on 30 retained tasks from internal DeepSeek R&D workloads, human-annotated rubrics; developer-reported | 5.4.4 Table 8 |
| R&D Coding Benchmark (internal) | Claude Opus 4.5 | 70% | pass rate on 30 retained tasks from internal DeepSeek R&D workloads, human-annotated rubrics; reported by DeepSeek in the DeepSeek-V4 technical report | 5.4.4 Table 8 |
| R&D Coding Benchmark (internal) | Claude Opus 4.5 Thinking | 73% | pass rate on 30 retained tasks from internal DeepSeek R&D workloads, human-annotated rubrics; reported by DeepSeek in the DeepSeek-V4 technical report | 5.4.4 Table 8 |
| R&D Coding Benchmark (internal) | Claude Opus 4.6 Thinking | 80% | pass rate on 30 retained tasks from internal DeepSeek R&D workloads, human-annotated rubrics; reported by DeepSeek in the DeepSeek-V4 technical report | 5.4.4 Table 8 |
| Internal developer survey: ready as default coding model | DeepSeek-V4-Pro | 52% | share of N=85 DeepSeek developers and researchers who said yes; developer-reported | 5.4.4 |
| Internal developer survey: ready as default coding model | DeepSeek-V4-Pro | 39% | share of N=85 DeepSeek developers and researchers who leaned toward yes; developer-reported | 5.4.4 |
| Internal developer survey: ready as default coding model | DeepSeek-V4-Pro | fewer than 9% | share of N=85 DeepSeek developers and researchers who said no; developer-reported | 5.4.4 |
| Chinese white-collar tasks (internal, human evaluation) | DeepSeek-V4-Pro-Max | 63% | non-loss rate vs. Opus-4.6-Max on 30 Chinese professional tasks, blind human annotation, in-house agent harness with Bash and web search; developer-reported | 5.4.3 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Easy — Objective Q&A | DeepSeek-V4-Pro | 56.1% | agentic search win rate, n=196 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Easy — Objective Q&A | DeepSeek-V4-Pro | 21.9% | RAG win rate, n=196 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Easy — Objective Q&A | DeepSeek-V4-Pro | 21.9% | tie rate, n=196 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Easy — Subjective Q&A | DeepSeek-V4-Pro | 61.7% | agentic search win rate, n=321 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Easy — Subjective Q&A | DeepSeek-V4-Pro | 17.4% | RAG win rate, n=321 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Easy — Subjective Q&A | DeepSeek-V4-Pro | 20.9% | tie rate, n=321 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Hard — Objective Q&A | DeepSeek-V4-Pro | 60.7% | agentic search win rate, n=168 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Hard — Objective Q&A | DeepSeek-V4-Pro | 19.6% | RAG win rate, n=168 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Hard — Objective Q&A | DeepSeek-V4-Pro | 19.6% | tie rate, n=168 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Hard — Subjective Q&A | DeepSeek-V4-Pro | 68.5% | agentic search win rate, n=184 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Hard — Subjective Q&A | DeepSeek-V4-Pro | 14.7% | RAG win rate, n=184 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Hard — Subjective Q&A | DeepSeek-V4-Pro | 16.8% | tie rate, n=184 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Total — all | DeepSeek-V4-Pro | 61.7% | agentic search win rate, n=869 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Total — all | DeepSeek-V4-Pro | 18.3% | RAG win rate, n=869 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Agentic Search vs. Retrieval Augmented Search (internal pairwise) — Total — all | DeepSeek-V4-Pro | 20.0% | tie rate, n=869 queries; agentic search vs. retrieval-augmented search for the same model; developer-reported | Appendix B Table 9 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Objective Q&A subtotal | DeepSeek-V4-Pro | 27.3% | DeepSeek-V4-Pro win rate, n=289; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Objective Q&A subtotal | DeepSeek-V4-Pro | 8.7% | DeepSeek-V3.2 win rate, n=289; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Objective Q&A subtotal | DeepSeek-V4-Pro | 64.0% | tie rate, n=289; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Subjective Q&A subtotal | DeepSeek-V4-Pro | 28.5% | DeepSeek-V4-Pro win rate, n=667; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Subjective Q&A subtotal | DeepSeek-V4-Pro | 11.1% | DeepSeek-V3.2 win rate, n=667; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Subjective Q&A subtotal | DeepSeek-V4-Pro | 60.4% | tie rate, n=667; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Total | DeepSeek-V4-Pro | 28.1% | DeepSeek-V4-Pro win rate, n=956; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Total | DeepSeek-V4-Pro | 10.4% | DeepSeek-V3.2 win rate, n=956; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Search Q&A pairwise vs. DeepSeek-V3.2 (internal, retrieval-augmented search) — Total | DeepSeek-V4-Pro | 61.5% | tie rate, n=956; internal pairwise evaluation; developer-reported | Appendix B Table 11 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Business Writing | DeepSeek-V4-Pro | 65.16% | DeepSeek-V4-Pro win rate, n=1349; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Business Writing | DeepSeek-V4-Pro | 32.32% | Gemini-3.1-Pro win rate, n=1349; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Business Writing | DeepSeek-V4-Pro | 2.52% | tie rate, n=1349; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Media Writing | DeepSeek-V4-Pro | 57.96% | DeepSeek-V4-Pro win rate, n=666; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Media Writing | DeepSeek-V4-Pro | 38.44% | Gemini-3.1-Pro win rate, n=666; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Media Writing | DeepSeek-V4-Pro | 3.60% | tie rate, n=666; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Everyday Writing | DeepSeek-V4-Pro | 69.49% | DeepSeek-V4-Pro win rate, n=390; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Everyday Writing | DeepSeek-V4-Pro | 25.90% | Gemini-3.1-Pro win rate, n=390; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Everyday Writing | DeepSeek-V4-Pro | 4.62% | tie rate, n=390; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Oral Writing | DeepSeek-V4-Pro | 58.62% | DeepSeek-V4-Pro win rate, n=319; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Oral Writing | DeepSeek-V4-Pro | 37.62% | Gemini-3.1-Pro win rate, n=319; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Oral Writing | DeepSeek-V4-Pro | 3.76% | tie rate, n=319; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Official Document | DeepSeek-V4-Pro | 54.78% | DeepSeek-V4-Pro win rate, n=230; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Official Document | DeepSeek-V4-Pro | 42.17% | Gemini-3.1-Pro win rate, n=230; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Official Document | DeepSeek-V4-Pro | 3.04% | tie rate, n=230; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Academic Writing | DeepSeek-V4-Pro | 63.43% | DeepSeek-V4-Pro win rate, n=216; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Academic Writing | DeepSeek-V4-Pro | 32.87% | Gemini-3.1-Pro win rate, n=216; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Academic Writing | DeepSeek-V4-Pro | 3.70% | tie rate, n=216; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Total | DeepSeek-V4-Pro | 62.65% | DeepSeek-V4-Pro win rate, n=3170; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Total | DeepSeek-V4-Pro | 34.10% | Gemini-3.1-Pro win rate, n=3170; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese functional writing pairwise vs. Gemini-3.1-Pro (internal) — Total | DeepSeek-V4-Pro | 3.25% | tie rate, n=3170; internal pairwise evaluation; developer-reported | Appendix B Table 12 |
| Chinese creative writing pairwise vs. Gemini-3.1-Pro (internal) — Instruction following — Total | DeepSeek-V4-Pro | 60.03% | DeepSeek-V4-Pro win rate, n=2837; internal pairwise evaluation; developer-reported | Appendix B Table 13 |
| Chinese creative writing pairwise vs. Gemini-3.1-Pro (internal) — Instruction following — Total | DeepSeek-V4-Pro | 39.44% | Gemini-3.1-Pro win rate, n=2837; internal pairwise evaluation; developer-reported | Appendix B Table 13 |
| Chinese creative writing pairwise vs. Gemini-3.1-Pro (internal) — Instruction following — Total | DeepSeek-V4-Pro | 0.53% | tie rate, n=2837; internal pairwise evaluation; developer-reported | Appendix B Table 13 |
| Chinese creative writing pairwise vs. Gemini-3.1-Pro (internal) — Writing quality — Total | DeepSeek-V4-Pro | 77.48% | DeepSeek-V4-Pro win rate, n=2837; internal pairwise evaluation; developer-reported | Appendix B Table 13 |
| Chinese creative writing pairwise vs. Gemini-3.1-Pro (internal) — Writing quality — Total | DeepSeek-V4-Pro | 22.35% | Gemini-3.1-Pro win rate, n=2837; internal pairwise evaluation; developer-reported | Appendix B Table 13 |
| Chinese creative writing pairwise vs. Gemini-3.1-Pro (internal) — Writing quality — Total | DeepSeek-V4-Pro | 0.18% | tie rate, n=2837; internal pairwise evaluation; developer-reported | Appendix B Table 13 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Complex instruction following | DeepSeek-V4-Pro | 46.9% | DeepSeek-V4-Pro win rate, n=49; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Complex instruction following | DeepSeek-V4-Pro | 53.1% | Claude Opus 4.5 win rate, n=49; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Complex instruction following | DeepSeek-V4-Pro | 0.0% | tie rate, n=49; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Multi-turn writing | DeepSeek-V4-Pro | 45.6% | DeepSeek-V4-Pro win rate, n=147; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Multi-turn writing | DeepSeek-V4-Pro | 51.7% | Claude Opus 4.5 win rate, n=147; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Multi-turn writing | DeepSeek-V4-Pro | 2.7% | tie rate, n=147; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Total | DeepSeek-V4-Pro | 45.9% | DeepSeek-V4-Pro win rate, n=196; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Total | DeepSeek-V4-Pro | 52.0% | Claude Opus 4.5 win rate, n=196; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
| Complex instruction following and multi-turn writing pairwise vs. Claude-Opus-4.5 (internal) — Total | DeepSeek-V4-Pro | 2.0% | tie rate, n=196; internal pairwise evaluation on the most challenging prompts; developer-reported | Appendix B Table 14 |
Provenance
The raw file is the arXiv HTML rendering of the report, scraped on September 19, 2026 and stored with the arXiv page chrome. All results are DeepSeek's own measurements, including those for other developers' models; the report does not describe safety, misuse or dangerous-capability evaluations. Confidence is medium because the figures are developer-reported and single-source.
Relationships
- supports: DeepSeek V4 Pro / V4 Flash — primary technical reference for the V4 preview models
- related: DeepSeek-V3 Technical Report — predecessor report whose MoE and multi-token-prediction designs V4 retains
- related: DeepSeek-V4.1-Flash Technical Report (DeepSeek, September 2026) — successor report
- instance-of: Mixture of Experts (MoE)