AIEvals

Models

DeepSeek V4.1 Flash 1017

The 17 October update to DeepSeek’s open-weight V4.1 Flash, with publisher-reported reasoning, coding, agent and visual results.

DeepSeekReleased 17 October 2026Context 1M tokensOutput 384k tokensOpen weights$0.30 in / $1.20 out per Mtok

Family DeepSeek V412 benchmarks with a published resultRead 17 October 2026

Where it stands

Rank compares tracked models within each publisher table. Published includes all reported entries in that table. Independent scores and ranks are not available.

BenchmarkScoreRankPublishedLeaderBehind
Text-only HLE (w/o tools)DeepSeek · publisher table · 17 October 202644.00%1of 31of 3DeepSeek V4.1 Flash 1017—
Apex Bench v1.1DeepSeek · publisher table · 17 October 202672.10%1of 61of 6DeepSeek V4.1 Flash 1017—
Terminal-Bench 4.0DeepSeek · publisher table · 17 October 202642.30%2of 62of 6Claude Opus 59.50 pts
ProgramBenchDeepSeek · publisher table · 17 October 202629.10%2of 62of 6Claude Opus 57.90 pts
AutomationBench (v1.0.6)DeepSeek · publisher table · 17 October 202660.50%1of 61of 6DeepSeek V4.1 Flash 1017—
DeepSWE v1.1DeepSeek · publisher table · 17 October 202677.00%1of 61of 6DeepSeek V4.1 Flash 1017—
Chartography (w/tools)DeepSeek · publisher table · 17 October 202683.00%2of 52of 5Claude Opus 51.00 pts
HLE (w/o tools)DeepSeek · publisher table · 17 October 202642.60%4of 54of 5Claude Opus 513.70 pts
Agents' Last ExamDeepSeek · publisher table · 17 October 202635.40%1of 61of 6DeepSeek V4.1 Flash 1017—
NL2Repo-BenchDeepSeek · publisher table · 17 October 202668.20%2of 62of 6Claude Opus 57.10 pts
BabyVision (w/tools)DeepSeek · publisher table · 17 October 202692.50%2of 52of 5Claude Opus 51.60 pts
ZeroBench-main (w/tools)DeepSeek · publisher table · 17 October 202651.00%3of 53of 5GPT-5.6 Sol2.00 pts

Results by benchmark

Independent and vendor results retain their own harnesses and effort settings.

BenchmarkIndependent runAs reported by a publisher
Humanity's Last Exam (text-only, without tools)Multidisciplinary reasoningno published result
44.0%Text-only HLE (w/o tools)DeepSeek · 17 October 2026Rank 1 of 3 in this publisher table
Apex Bench v1.1Reasoningno published result
72.1%Apex Bench v1.1DeepSeek · 17 October 2026Rank 1 of 6 in this publisher table
Terminal-Bench 4.0Agentic codingno published result
42.3%Terminal-Bench 4.0DeepSeek · 17 October 2026Rank 2 of 6 in this publisher table
ProgramBenchAgentic codingno published result
29.1%ProgramBenchDeepSeek · 17 October 2026Rank 2 of 6 in this publisher table
AutomationBenchBusiness workflowsno published result
60.5%AutomationBench (v1.0.6)DeepSeek · 17 October 2026Rank 1 of 6 in this publisher table
DeepSWE v1.1Software engineeringno published result
77.0%DeepSWE v1.1DeepSeek · 17 October 2026Rank 1 of 6 in this publisher table
ChartographyVisual chart recognitionno published result
83.0%Chartography (w/tools)DeepSeek · 17 October 2026Rank 2 of 5 in this publisher table
Humanity's Last ExamMultidisciplinary reasoningno published result
42.6%HLE (w/o tools)DeepSeek · 17 October 2026Rank 4 of 5 in this publisher table
Agents' Last ExamComputer useno published result
35.4%Agents' Last ExamDeepSeek · 17 October 2026Rank 1 of 6 in this publisher table
NL2Repo-BenchAgentic codingno published result
68.2%NL2Repo-BenchDeepSeek · 17 October 2026Rank 2 of 6 in this publisher table
BabyVisionMultimodalno published result
92.5%BabyVision (w/tools)DeepSeek · 17 October 2026Rank 2 of 5 in this publisher table
ZeroBench (main set)Multimodalno published result
51.0%ZeroBench-main (w/tools)DeepSeek · 17 October 2026Rank 3 of 5 in this publisher table
Independent column: the suite named under each figure, one harness per suite. Publisher column: each lab’s own table, reproduced on the benchmark’s page. Publisher ranks compare reported scores in the same table row only; ties share a rank and unreported cells are excluded.

Source notes

From the linked sources. Read 17 October 2026.

Release and API alias. Released 17 October 2026 on the official API, web and app. Use deepseek-flash to call DeepSeek-V4.1-Flash-1017; the API alias is not an immutable snapshot identifier. The September V4.1 Flash record and its historical results remain separate.

Context, output and capabilities. 1M-token context and up to 384K output tokens. Text and image input, text output; thinking (default) and non-thinking modes, tool calls, JSON output, Responses API and Anthropic-compatible API are supported.

Peak and off-peak pricing. USD per million tokens: peak input $0.30, cached input $0.006, output $1.20; off-peak input $0.15, cached input $0.003, output $0.60. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays; all other hours use off-peak rates. The headline shows peak cache-miss input/output rates. Prices are unchanged by the 1017 update.

Weights and local inference. The dated model repository and weights are released under the MIT License. Its reference encoder replaces a Jinja chat template; deepseek-recipe provides protocol-aware prompt encoding. The model card recommends temperature 1.0, top_p 1.0, a 1M context window and a generation budget of at least 256K tokens.

Evaluation settings and subsets. The complete 11-row, six-model comparison is retained. HLE without tools is 42.6% on the full set and 44.0% on the text-only subset; subset scores have their own benchmark page and ranks. Terminal-Bench 4.0 uses Claude Code, DeepSWE v1.1 uses Mini-SWE-V2, and ZeroBench-main uses tools and Pass@5. Automation-Bench v1.0.6 and Apex Bench v1.1 retain their versions; Apex Bench is distinct from MathArena Apex. The release does not specify all effort levels, trial counts or judge settings.

Independent coverage. No separately identified 1017 result was found in the Vals AI or Artificial Analysis public model directories or Datacurve’s DeepSWE v1.1 public artifact checked on 17 October 2026. Earlier V4.1 Flash scores are not assigned to this update. All results here are publisher reports and do not fill independent comparison columns.

Efficiency and harness compatibility. DeepSeek claims shorter task runtimes and outputs, with optimizations for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. These are publisher claims rather than measured API speed or latency. DeepSeek recommends DeepSeek Harness 0.2.1-rc.1 or later.

Sources

Read on 17 October 2026.