"DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression" is the technical report DeepSeek published with the weights of DeepSeek-V4.1-Flash on Hugging Face (Source: huggingface.co). The PDF carries no date; the model was released on September 9, 2026, and the report refers to a production deployment "launched in September 2026." The repository is MIT-licensed. The report describes a multimodal mixture-of-experts model with 552B backbone parameters that accepts images and text and supports contexts of up to one million tokens.
Architecture and training
The report's stated aim is to reduce the cost of input-heavy agentic workloads by shrinking the key-value cache. Its main elements are a Causal Encoder-Decoder (CED) architecture, a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder, which activates 8B parameters per token during prefill and 16B during decode; Compressed Sparse Attention 2 (CSA2), which shares key-value entries and sparse-attention indices across layers; FP4 caching of the main KV entries; and SWA Bounded Replay, a deployment technique that reconstructs sliding-window attention state by replaying recent tokens. DeepSeek states that these reduce the global KV cache to 890 bytes per token, about one quarter of DeepSeek-V4-Flash's, and the persistent KV cache to about one eighth. Other components include Single-Pass mHC, a 196B-parameter Engram conditional memory, DSpark speculative decoding, and a vision encoder trained from scratch.
The model was pre-trained from scratch on a multimodal corpus of 45T tokens. Post-training follows supervised fine-tuning, reinforcement learning and on-policy distillation; the report says the substantive changes are in the data pipeline, which synthesizes agent tasks and environments at scale. Reasoning effort is a scalar from 1 to 100, and the public API's low, high and max tiers correspond to 50, 75 and 100.
Evaluation approach
Base-model results (Table 1) compare V4.1-Flash-Base with V4-Flash-Base and V4-Pro-Base in DeepSeek's internal framework across world knowledge, language and reasoning, code and mathematics, long context, and four multimodal benchmarks. The report also gives bits-per-byte on internal held-out corpora as a chart.
Post-training evaluation covers reasoning (GPQA Diamond, Humanity's Last Exam, an internal Codeforces benchmark, MathArena Apex) and four agentic categories: code agents (Terminal-Bench 2.1, 3.0 and 4.0, DeepSWE v1.1, ProgramBench, NL2Repo-Bench), cybersecurity (SEC-Bench Pro, CyberGym, ExploitGym), general agents (AutomationBench v1.0.6 public set, Agents' Last Exam), and visual agents (Chartography, BabyVision, ZeroBench). All Table 3 results use maximum reasoning effort. Code-agent runs use the Minimal mode of DeepSeek Harness with a one-million-token context, except DeepSWE v1.1, which uses the mini-SWE harness, and SEC-Bench Pro and the visual tasks, which use Claude Code. Table 3 compares the model with Claude Opus 5, GPT-5.6 Sol, Kimi K3, GLM-5.3 and the two DeepSeek-V4 models, without naming the V4 checkpoints or the origin of the other developers' figures. Table 4 repeats DeepSWE v1.1 and Terminal-Bench 2.1 under eight scaffold configurations, and an appendix table under four Claude Code versions.
DeepSeek says it restricted internet access, stripped Git histories and purged build caches in coding-agent environments to limit reward hacking, and that it still observed exploit-seeking behaviour, giving the example of the model decompiling Ubuntu packages to find vulnerabilities in CyberGym.
Findings as stated by DeepSeek
The report presents V4.1-Flash as improving on V4-Flash across reasoning and agentic benchmarks and as matching or exceeding the compared open and proprietary models on several agentic tasks, including DeepSWE v1.1, Terminal-Bench 2.1, AutomationBench and Agents' Last Exam. It states that the model sets the highest score among open-source models on the cybersecurity tasks and, citing the dual-use nature of those capabilities, encourages defensive use. On visual-agent tasks it reports scores above Kimi K3 and below the closed models. Accuracy and output length rise with the reasoning-effort setting, with most of the gain reached by effort 60 to 80. Preliminary multi-agent experiments with DeepSeek Harness's Agent Team mode report higher scores than single-agent runs at every wall-clock deadline on ProgramBench and FrontierSWE v2.
The conclusion states that the new architecture creates "robustness boundaries that have yet to be fully characterized," naming possible selection errors in CSA2 and approximate state reconstruction in SWA Bounded Replay. It also says that narrow benchmark margins do not mean the model matches leading closed systems on the hardest tasks, naming Fable-5 and GPT-6 Astra.
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-V4-Flash-Base | 83.9 | base model; 3–5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| AGIEval | DeepSeek-V4-Pro-Base | 84.4 | base model; 3–5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| AGIEval | DeepSeek-V4.1-Flash-Base | 83.4 | base model; 3–5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 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; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 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; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| C-Eval | DeepSeek-V4.1-Flash-Base | 92.1 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| MultiLoKo | DeepSeek-V4-Flash-Base | 42.6 | base model; 5-shot; LLM-Judge; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| MultiLoKo | DeepSeek-V4-Pro-Base | 50.9 | base model; 5-shot; LLM-Judge; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| MultiLoKo | DeepSeek-V4.1-Flash-Base | 45.5 | base model; 5-shot; LLM-Judge; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| SuperGPQA | DeepSeek-V4-Flash-Base | 46.5 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| SuperGPQA | DeepSeek-V4-Pro-Base | 53.9 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| SuperGPQA | DeepSeek-V4.1-Flash-Base | 53.1 | base model; 5-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| BBH | DeepSeek-V4-Flash-Base | 86.9 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| BBH | DeepSeek-V4-Pro-Base | 87.5 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| BBH | DeepSeek-V4.1-Flash-Base | 86.1 | base model; 3-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| BBEH | DeepSeek-V4-Flash-Base | 25.4 | base model; 1-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| BBEH | DeepSeek-V4-Pro-Base | 29.8 | base model; 1-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| BBEH | DeepSeek-V4.1-Flash-Base | 27.2 | base model; 1-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| DROP | DeepSeek-V4-Flash-Base | 88.6 | base model; 1-shot; F1; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| DROP | DeepSeek-V4-Pro-Base | 88.7 | base model; 1-shot; F1; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| DROP | DeepSeek-V4.1-Flash-Base | 87.9 | base model; 1-shot; F1; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| HellaSwag | DeepSeek-V4-Flash-Base | 85.7 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| HellaSwag | DeepSeek-V4-Pro-Base | 88.0 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| HellaSwag | DeepSeek-V4.1-Flash-Base | 87.2 | base model; 0-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 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; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 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; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| BigCodeBench | DeepSeek-V4.1-Flash-Base | 60.6 | base model; 3-shot; Pass@1; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| MGSM | DeepSeek-V4-Flash-Base | 85.7 | base model; 8-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| MGSM | DeepSeek-V4-Pro-Base | 84.4 | base model; 8-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| MGSM | DeepSeek-V4.1-Flash-Base | 80.2 | base model; 8-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 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; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 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; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report (DeepSeek's own earlier base model, same internal framework) | 4.3.2 Table 1 |
| LongBench-V2 | DeepSeek-V4.1-Flash-Base | 45.2 | base model; 1-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| CVBench | DeepSeek-V4.1-Flash-Base | 77.9 | base model; 4-shot; EM; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| DocVQA | DeepSeek-V4.1-Flash-Base | 95.6 | base model; 4-shot; LLM-Judge; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| RefCOCO-avg | DeepSeek-V4.1-Flash-Base | 86.0 | base model; 0-shot; Acc@0.5; DeepSeek internal evaluation framework, same setting for all three models; developer-reported | 4.3.2 Table 1 |
| DeepSWE v1.1 | Claude Opus 5 | 74.0 | Resolved; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| DeepSWE v1.1 | GPT-5.6 Sol | 73.0 | Resolved; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| DeepSWE v1.1 | Kimi K3 | 67.5 | Resolved; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| DeepSWE v1.1 | GLM-5.3 | 66.9 | Resolved; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| DeepSWE v1.1 | DeepSeek-V4-Pro | 62.7 | Resolved; Max reasoning effort; mini-SWE harness (official setup), 1M-token context, temperature 1.0, top-p 0.95; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| DeepSWE v1.1 | DeepSeek-V4-Flash | 54.4 | Resolved; Max reasoning effort; mini-SWE harness (official setup), 1M-token context, temperature 1.0, top-p 0.95; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 74.2 | Resolved; Max reasoning effort (reasoning_effort=100); mini-SWE harness (official setup), 1M-token context, temperature 1.0, top-p 0.95; developer-reported | 5.3.2 Table 3 |
| ProgramBench | Claude Opus 5 | 37.0 | Almost@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| ProgramBench | GPT-5.6 Sol | 23.0 | Almost@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| ProgramBench | Kimi K3 | 17.5 | Almost@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| ProgramBench | GLM-5.3 | 19.0 | Almost@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| ProgramBench | DeepSeek-V4-Pro | 15.5 | Almost@1; Max reasoning effort; DeepSeek Harness Minimal mode, 1M-token context, temperature 1.0, top-p 0.95, internet access restricted and Git histories stripped; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| ProgramBench | DeepSeek-V4.1-Flash | 20.3 | Almost@1; Max reasoning effort (reasoning_effort=100); DeepSeek Harness Minimal mode, 1M-token context, temperature 1.0, top-p 0.95, internet access restricted and Git histories stripped; developer-reported | 5.3.2 Table 3 |
| NL2Repo-Bench | Claude Opus 5 | 75.3 | Score; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| NL2Repo-Bench | GPT-5.6 Sol | 56.8 | Score; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| NL2Repo-Bench | Kimi K3 | 58.0 | Score; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| NL2Repo-Bench | GLM-5.3 | 58.0 | Score; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| NL2Repo-Bench | DeepSeek-V4-Pro | 61.5 | Score; Max reasoning effort; DeepSeek Harness Minimal mode, 1M-token context, temperature 1.0, top-p 0.95, internet access restricted and Git histories stripped; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| NL2Repo-Bench | DeepSeek-V4-Flash | 54.2 | Score; Max reasoning effort; DeepSeek Harness Minimal mode, 1M-token context, temperature 1.0, top-p 0.95, internet access restricted and Git histories stripped; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| NL2Repo-Bench | DeepSeek-V4.1-Flash | 65.4 | Score; Max reasoning effort (reasoning_effort=100); DeepSeek Harness Minimal mode, 1M-token context, temperature 1.0, top-p 0.95, internet access restricted and Git histories stripped; developer-reported | 5.3.2 Table 3 |
| SEC-Bench Pro | GPT-5.6 Sol | 74.3 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| SEC-Bench Pro | DeepSeek-V4-Pro | 56.4 | Pass@1; Max reasoning effort; SEC-Bench Pro version 260505, Claude Code harness; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| SEC-Bench Pro | DeepSeek-V4-Flash | 30.9 | Pass@1; Max reasoning effort; SEC-Bench Pro version 260505, Claude Code harness; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| SEC-Bench Pro | DeepSeek-V4.1-Flash | 62.8 | Pass@1; Max reasoning effort (reasoning_effort=100); SEC-Bench Pro version 260505, Claude Code harness; developer-reported | 5.3.2 Table 3 |
| AutomationBench | Claude Opus 5 | 50.3 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| AutomationBench | GPT-5.6 Sol | 45.8 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| AutomationBench | Kimi K3 | 46.7 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| AutomationBench | GLM-5.3 | 48.8 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| AutomationBench | DeepSeek-V4-Pro | 43.2 | Pass@1; Max reasoning effort; public evaluation set of AutomationBench v1.0.6, official scaffold; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| AutomationBench | DeepSeek-V4-Flash | 37.7 | Pass@1; Max reasoning effort; public evaluation set of AutomationBench v1.0.6, official scaffold; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| AutomationBench | DeepSeek-V4.1-Flash | 54.8 | Pass@1; Max reasoning effort (reasoning_effort=100); public evaluation set of AutomationBench v1.0.6, official scaffold; developer-reported | 5.3.2 Table 3 |
| Agent's Last Exam | Claude Opus 5 | 28.6 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| Agent's Last Exam | GPT-5.6 Sol | 26.7 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| Agent's Last Exam | Kimi K3 | 27.6 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| Agent's Last Exam | GLM-5.3 | 28.5 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| Agent's Last Exam | DeepSeek-V4-Pro | 25.7 | Pass@1; Max reasoning effort; ALE-CLI, official scaffold; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| Agent's Last Exam | DeepSeek-V4-Flash | 25.2 | Pass@1; Max reasoning effort; ALE-CLI, official scaffold; developer-reported; checkpoint of the V4 model not stated in the report | 5.3.2 Table 3 |
| Agent's Last Exam | DeepSeek-V4.1-Flash | 31.8 | Pass@1; Max reasoning effort (reasoning_effort=100); ALE-CLI, official scaffold; developer-reported | 5.3.2 Table 3 |
| Chartography w/ tools | Claude Opus 5 | 84.0 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| Chartography w/ tools | GPT-5.6 Sol | 79.9 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| Chartography w/ tools | Kimi K3 | 68.1 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| Chartography w/ tools | DeepSeek-V4.1-Flash | 78.9 | Pass@1; Max reasoning effort (reasoning_effort=100); Claude Code harness, 512k-token context, temperature 1.0, top-p 0.95; developer-reported | 5.3.2 Table 3 |
| BabyVision w/ tools | Claude Opus 5 | 94.1 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| BabyVision w/ tools | GPT-5.6 Sol | 88.9 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| BabyVision w/ tools | Kimi K3 | 85.7 | Pass@1; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| BabyVision w/ tools | DeepSeek-V4.1-Flash | 89.6 | Pass@1; Max reasoning effort (reasoning_effort=100); Claude Code harness, 512k-token context, temperature 1.0, top-p 0.95; developer-reported | 5.3.2 Table 3 |
| ZeroBench-main w/ tools | Claude Opus 5 | 52.0 | Pass@5; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| ZeroBench-main w/ tools | GPT-5.6 Sol | 53.0 | Pass@5; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| ZeroBench-main w/ tools | Kimi K3 | 41.0 | Pass@5; Max reasoning effort; reported by DeepSeek in the DeepSeek-V4.1-Flash technical report; source of comparison figure not stated | 5.3.2 Table 3 |
| ZeroBench-main w/ tools | DeepSeek-V4.1-Flash | 49.0 | Pass@5; Max reasoning effort (reasoning_effort=100); Claude Code harness, 512k-token context, temperature 1.0, top-p 0.95; developer-reported | 5.3.2 Table 3 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 69.8 | Resolved; scaffold: Claude Code v2.1.251; Max reasoning effort (100); N=8 samples per task; Linux containers, temperature 1.0, top-p 0.95, 1M-token context, max_steps=500; developer-reported | 5.3.4 Table 4 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 65.6 | Resolved; scaffold: Codex v0.147.0; Max reasoning effort (100); N=8 samples per task; Linux containers, temperature 1.0, top-p 0.95, 1M-token context, max_steps=500; developer-reported | 5.3.4 Table 4 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 65.5 | Resolved; scaffold: OpenCode v1.18.15; Max reasoning effort (100); N=8 samples per task; Linux containers, temperature 1.0, top-p 0.95, 1M-token context, max_steps=500; developer-reported | 5.3.4 Table 4 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 66.2 | Resolved; scaffold: Pi v0.84.2; Max reasoning effort (100); N=8 samples per task; Linux containers, temperature 1.0, top-p 0.95, 1M-token context, max_steps=500; developer-reported | 5.3.4 Table 4 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 72.6 | Resolved; scaffold: DeepSeek Harness Minimal; Max reasoning effort (100); N=8 samples per task; Linux containers, temperature 1.0, top-p 0.95, 1M-token context, max_steps=500; developer-reported | 5.3.4 Table 4 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 70.5 | Resolved; scaffold: DeepSeek Harness Standard; Max reasoning effort (100); N=8 samples per task; Linux containers, temperature 1.0, top-p 0.95, 1M-token context, max_steps=500; developer-reported | 5.3.4 Table 4 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 67.6 | Resolved; scaffold: DeepSeek Harness PTC; Max reasoning effort (100); N=8 samples per task; Linux containers, temperature 1.0, top-p 0.95, 1M-token context, max_steps=500; developer-reported | 5.3.4 Table 4 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 68.4 | scaffold: Claude Code v2.1.105; Max reasoning effort; settings as Table 4; developer-reported | Appendix B.1 Table 5 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 68.7 | scaffold: Claude Code v2.1.238; Max reasoning effort; settings as Table 4; developer-reported | Appendix B.1 Table 5 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 68.6 | scaffold: Claude Code v2.1.259; Max reasoning effort; settings as Table 4; developer-reported | Appendix B.1 Table 5 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 68.9 | scaffold: Claude Code average of four versions (from unrounded scores); Max reasoning effort; settings as Table 4; developer-reported | Appendix B.1 Table 5 |
| Average Pass@1 over eight reasoning benchmarks | DeepSeek-V4.1-Flash | 67.1% | AIME 2026, Apex 2025 Shortlist, GPQA Diamond, HLE, IMO-AnswerBench, LiveCodeBench, MathArena-Apex, SimpleQA-Verified; reasoning effort 25; developer-reported | 5.3.2 |
| Average Pass@1 over eight reasoning benchmarks | DeepSeek-V4.1-Flash | 76.3% | AIME 2026, Apex 2025 Shortlist, GPQA Diamond, HLE, IMO-AnswerBench, LiveCodeBench, MathArena-Apex, SimpleQA-Verified; reasoning effort 100; developer-reported | 5.3.2 |
| DeepSWE v1.1 | DeepSeek-V4.1-Flash | 66.0% | mini-SWE; reasoning effort 25; developer-reported | 5.3.2 |
| ProgramBench (172 'golden' tasks) | DeepSeek-V4.1-Flash | 13.59% | Almost@1 (fraction of rollouts scoring at least 0.95), up to three rollouts per task; multi-agent (DeepSeek Harness Agent Team); per-rollout wall-clock deadline 1 hour; described as preliminary; developer-reported | 5.3.5 |
| ProgramBench (172 'golden' tasks) | DeepSeek-V4.1-Flash | 30.04% | Almost@1 (fraction of rollouts scoring at least 0.95), up to three rollouts per task; multi-agent (DeepSeek Harness Agent Team); per-rollout wall-clock deadline 8 hours; described as preliminary; developer-reported | 5.3.5 |
| ProgramBench (172 'golden' tasks) | DeepSeek-V4.1-Flash | 12.79% | Almost@1 (fraction of rollouts scoring at least 0.95), up to three rollouts per task; single-agent; per-rollout wall-clock deadline 1 hour; described as preliminary; developer-reported | 5.3.5 |
| ProgramBench (172 'golden' tasks) | DeepSeek-V4.1-Flash | 20.39% | Almost@1 (fraction of rollouts scoring at least 0.95), up to three rollouts per task; single-agent; per-rollout wall-clock deadline 8 hours; described as preliminary; developer-reported | 5.3.5 |
| FrontierSWE v2 (no-GPU subset) | DeepSeek-V4.1-Flash | 13.50% | Mean@5; multi-agent (DeepSeek Harness Agent Team); per-rollout wall-clock deadline 1 hour; described as preliminary; developer-reported | 5.3.5 |
| FrontierSWE v2 (no-GPU subset) | DeepSeek-V4.1-Flash | 32.90% | Mean@5; multi-agent (DeepSeek Harness Agent Team); per-rollout wall-clock deadline 20 hours; described as preliminary; developer-reported | 5.3.5 |
| FrontierSWE v2 (no-GPU subset) | DeepSeek-V4.1-Flash | 10.50% | Mean@5; single-agent; per-rollout wall-clock deadline 1 hour; described as preliminary; developer-reported | 5.3.5 |
| FrontierSWE v2 (no-GPU subset) | DeepSeek-V4.1-Flash | 28.20% | Mean@5; single-agent; per-rollout wall-clock deadline 20 hours; described as preliminary; developer-reported | 5.3.5 |
Provenance
The raw file is a text parse of the PDF in the model's Hugging Face repository, captured on September 19, 2026. Tables 1 and 3 parse with damaged headers; the model card carries clean copies, which agree with the report except for one cell (NL2Repo-Bench for V4.1-Flash: 65.4 in the report, 64.0 on the model card). All results are DeepSeek's own. The report contains no safety, refusal or dangerous-capability evaluation beyond the cybersecurity benchmark scores. Confidence is medium because the figures are developer-reported and single-source.
Relationships
- supports: DeepSeek-V4.1-Flash — primary technical reference
- depends-on: DeepSeek-V4 Technical Report (DeepSeek, April 2026) — the architecture it revises
- related: Inference Economics and Token Pricing — the KV cache and activation claims
- instance-of: Mixture of Experts (MoE)