Inference
EN

Hy4 preview vs Kimi K3

Change models
Hy4 previewTencent
Kimi K3Moonshot AI
IndexAll categories, out of 100.
66.4 ±4.1
66.9 ±2.4 best
GapIndex points below the top.
−0.5 · no clear gap
Top
CoverageIndex weight that has results.
90%
95% best
SpeedOutput tokens per second.
67/s best
44/s
EloLMArena rating and rank.
—
1474
ReleasedMonth of release.
2026-08
2026-07
Input / 1MUS dollars per 1M input tokens.
$0.834
$0.192 best
Output / 1MUS dollars per 1M output tokens.
$2.50 best
$16.00
Batch output / 1MBatch price per 1M output tokens.
—
$11.40
1K chats2K tokens in, 500 out, each.
$2.92 best
$8.38
1K documents20K tokens in, 1K out, each.
$19.18 best
$19.84
1K agent tasks200K tokens in, 20K out, each.
$216.82 best
$358.40
ContextMaximum tokens in one request.
1.05M
1.05M
AgenticMulti-step tasks with tools.
63.5
65.9 best
CodingCode writing and repair.
65.5
66.2 best
ReasoningLogic problems and puzzles.
69.6 best
67.7
MultimodalTasks with images and text.
75.7 best
65.6
KnowledgeFacts and expert knowledge.
67.9 best
66.3
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
—
71.0
MathMath problems.
73.0 best
66.6

Benchmarks

40 shared
Only Hy4 preview: 13Only Kimi K3: 89
Hy4 preview
Kimi K3
Critical Physics TasksReasoning
16.9%43 / 214
23.4% best23 / 214
LegalBenchLegal
83.8%47 / 149
86.2% best15 / 149
Graduate-Level Google-Proof Q&AKnowledge
92.3%18 / 141
93.5% best10 / 141
Vibe Code Bench v1.1Coding
77.5%37 / 123
85.0% best17 / 123
MedScribeHealthcare
83.6%45 / 109
88.0% best15 / 109
MedCodeHealthcare
43.2%47 / 106
49.4% best23 / 106
Legal Research BenchLegal
45.2%22 / 90
46.2% best19 / 90
Terminal-Bench 2.1 · BenchLMAgentic
85.4%18 / 86
88.3% best9 / 86
GPQA Diamond · BenchLMKnowledge
92.3%19 / 84
93.5% best10 / 84
Tax Agent BenchFinance
32.6%21 / 82
33.4% best19 / 82
Finance Agent (v2)Finance
55.1% best23 / 79
53.1%35 / 79
Code MigrationCoding
47.4% best14 / 77
16.1%55 / 77
Humanity's Last ExamKnowledge
55.4%14 / 77
56.0% best13 / 77
EMBFinance
58.1%38 / 75
66.7% best21 / 75
Terminal-Bench 2.1 · Vals AIAgentic
55.1%49 / 75
80.9% best9 / 75
ProgramBenchCoding
0.0%34 / 65
2.0% best11 / 65
Terminal-Bench 4.0Agentic
8.1%48 / 61
17.2% best37 / 61
DeepSWEAgentic
64.3%36 / 59
67.5% best28 / 59
Agent Arena steerabilityAgentic
0.2 best26 / 54
−0.828 / 54
Agent Arena task outcomeAgentic
3.5 best21 / 54
3.422 / 54