Inference
EN

DeepSeek V4.1 Flash 1017 vs Hy4

Change models
DeepSeek V4.1 Flash 1017DeepSeekFastest
Hy4Tencent
IndexAll categories, out of 100.
68.5 ±2.6
72.0 ±7.8 best
GapIndex points below the top.
−3.5 · no clear gap
Top
CoverageIndex weight that has results.
95% best
65%
SpeedOutput tokens per second.
261/s best
52/s
EloLMArena rating and rank.
1463
—
ReleasedMonth of release.
2026-10
2026-11
Input / 1MUS dollars per 1M input tokens.
$0.022 best
$0.417
Output / 1MUS dollars per 1M output tokens.
$0.40 best
$1.25
Batch output / 1MBatch price per 1M output tokens.
$0.20
—
1K chats2K tokens in, 500 out, each.
$0.244 best
$1.46
1K documents20K tokens in, 1K out, each.
$0.84 best
$9.59
1K agent tasks200K tokens in, 20K out, each.
$12.40 best
$108.41
ContextMaximum tokens in one request.
1.05M
1.05M
AgenticMulti-step tasks with tools.
72.2
74.5 best
CodingCode writing and repair.
69.8 best
68.4
ReasoningLogic problems and puzzles.
65.0
—
MultimodalTasks with images and text.
62.2 best
57.4
KnowledgeFacts and expert knowledge.
64.2
83.7 best
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
68.3
—
MathMath problems.
65.4
—

Benchmarks

14 shared
Only DeepSeek V4.1 Flash 1017: 80Only Hy4: 7
DeepSeek V4.1 Flash 1017
Hy4
Vibe Code Bench v1.1Coding
84.7% best18 / 123
81.0%30 / 123
Legal Research BenchLegal
41.3%30 / 90
41.3%30 / 90
Tax Agent BenchFinance
26.6%41 / 82
34.4% best17 / 82
Humanity's Last ExamKnowledge
42.6%31 / 77
59.1% best7 / 77
Terminal-Bench 4.0Agentic
19.7%34 / 61
39.4% best15 / 61
DeepSWEAgentic
77.0% best4 / 59
74.9%8 / 59
Agent Arena steerabilityAgentic
4.018 / 54
8.8 best14 / 54
Agent Arena task outcomeAgentic
2.325 / 54
6.7 best12 / 54
Humanity's Last Exam with toolsAgentic
63.9% best5 / 36
59.1%10 / 36
AutomationBenchAgentic
60.5% best3 / 32
56.2%4 / 32
FrontierCode 2.0Coding
19.3%23 / 27
31.0% best9 / 27
Agents' Last ExamAgentic
35.4%11 / 21
41.0% best8 / 21
ProgramBench: Can Language Models Rebuild Programs From Scratch?Coding
29.1%14 / 19
30.0% best12 / 19
ZeroBench_main with PythonMultimodal
51.0%
55.4%