BlogLLM Benchmark Scores as Published, Side by Side

DeepSeek-V4-Flash (Preview): Benchmark Scores as Published

Every result the catalog holds for DeepSeek-V4-Flash (Preview): 53 tests in 11 areas, from 4 pages. 31 printed by DeepSeek itself, the rest by the company named beside each number. Each number is as printed; nothing is ranked or rescored.

Developer DeepSeekReleased 2026-04-24Context window 1MInput TextOutput TextWeights Open (MIT)Data version 2026103001, as of 2026-10-30

Knowledge Work

Agentic Coding

TestValuePrinted bySetup as printed
SWE-bench Verified79%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15; printed on 1 more pageEffort: max
73.7%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
78.6%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
SWE-Bench Pro52.6%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
49.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
52.3%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
Terminal-Bench 2.161.8%DeepSeek, DeepSeek-V4-Flash-0731 model cardchecked 2026-10-15; printed on 2 more pages-
DeepSWE v1.17.3%DeepSeek, DeepSeek-V4-Flash-0731 model cardchecked 2026-10-15; printed on 1 more page-
SWE-bench Multilingual73.3%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15; printed on 1 more pageEffort: max
70.2%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
69.7%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
Terminal-Bench 2.056.6%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
56.9%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
49.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
NL2Repo-Bench39.4DeepSeek, DeepSeek-V4-Flash-0731 model cardchecked 2026-10-15; printed on 1 more page-
DSBench-FullStack37%DeepSeek, DeepSeek-V4-Flash-0731 model cardchecked 2026-10-15; printed on 1 more page-
DSBench-Hard25.8%DeepSeek, DeepSeek-V4-Flash-0731 model cardchecked 2026-10-15; printed on 1 more page-

Science and Math

TestValuePrinted bySetup as printed
GPQA Diamond87.4%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
88.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
71.2%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
SciCode48.2%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15-
IMO-AnswerBench88.4%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15; printed on 1 more pageEffort: max
41.9%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
85.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
HMMT February 202640.8%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
91.9%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
94.8%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
CritPt10.6%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15Tools: none
MathArena Apex Shortlist9.3%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15; printed on 1 more pageEffort: non-think
72.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
85.7%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
MathArena Apex1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
19.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
33%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
GPQA88.5%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15Tools: none

Long Context

TestValuePrinted bySetup as printed
AA-LCR62.7%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15-
LongBench v257%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15-
CorpusQA 1M59.3%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high; Context: 1M
15.5%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think; Context: 1M
60.5%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max; Context: 1M
MRCR 1M (MMR)76.9%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high; Context: 1M
78.7%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max; Context: 1M
37.5%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think; Context: 1M
RULER at 1M87.7%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15Context: 1048576

Cybersecurity

TestValuePrinted bySetup as printed
CyberGym38.7%DeepSeek, DeepSeek-V4-Flash-0731 model cardchecked 2026-10-15; printed on 1 more page-

General Knowledge and Reasoning

TestValuePrinted bySetup as printed
Humanity's Last Exam (no tools)34.8%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15; printed on 2 more pagesEffort: max; Tools: none
29.4%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high; Tools: none
8.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think; Tools: none
MMLU-Pro86.2%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15; printed on 1 more pageEffort: max
86.4%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
83%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
Humanity's Last Exam (with tools)45.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15; printed on 2 more pagesEffort: max; Tools: enabled; exact tools in source
40.3%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high; Tools: enabled; exact tools in source
AA-Omniscience Accuracy39.9%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15-
SimpleQA Verified28.9%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
34.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
23.1%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
AA-Omniscience Non-Hallucination2.8%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15-
Chinese SimpleQA78.9%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
71.5%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
73.2%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high

Code Generation

TestValuePrinted bySetup as printed
LiveCodeBench v690.9%NVIDIA, NVIDIA Nemotron 3 Ultra 550B-A55B model cardchecked 2026-10-15-
Codeforces rating2816DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high
3052DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
LiveCodeBench (version not stated)91.6%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: max
55.2%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: non-think
88.4%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high

Tool Use and Workflows

Web Search and Research

TestValuePrinted bySetup as printed
BrowseComp73.2%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15; printed on 1 more pageEffort: max; Tools: enabled; exact tools in source
53.5%DeepSeek, DeepSeek-V4-Pro (preview) model cardchecked 2026-10-15Effort: high; Tools: enabled; exact tools in source

Instruction Following

Multilingual

More from DeepSeek

Every number here is read from the same catalog file as the benchmark table, data version 2026103001. Where one page printed the model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.