Inference
EN
Alibaba

Qwen3.8 Max

Qwen3.8 Max is a reasoning model from Alibaba. 64 benchmarks count toward its index, in 7 categories.

  • Open weights
  • Reasoning
  • Input: text
  • Released 2026-08
RankedEnough results for an index.
IndexAll categories, out of 100.65.6±2.5Median 51.0
CoverageIndex weight that has results.95%
SpeedOutput tokens per second.41/sMedian 67/s
First tokenSeconds to the first output token.52.2 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.—Median $0.60
Output / 1MUS dollars per 1M output tokens.—Median $2.55
ContextMaximum tokens in one request.1MMedian 262K
EloLMArena rating and rank.1485±5 #16Median 1422

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.19 counted63.750 / 218
CodingCode writing and repair.18 counted64.347 / 209
ReasoningLogic problems and puzzles.5 counted66.739 / 227
MultimodalTasks with images and text.10 counted67.218 / 84
KnowledgeFacts and expert knowledge.5 counted62.652 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.2 counted71.42 / 63
MathMath problems.5 counted67.625 / 168

Results

64 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
CharXiv ReasoningMultimodal93.5%1 / 49
OSWorld-VerifiedAgentic86.1%1 / 37
MathVisionMultimodal95.2%1 / 24
GPQA Diamond · Vals AIKnowledge93.7% ±1.26 / 135
IOI v1Coding73.0% ±1.03 / 61
Instruction Following BenchmarkInstruction82.8%3 / 57
ERQAMultimodal77.8%1 / 18
LiveCodeBenchCoding87.9% ±1.010 / 142
τ²-bench BankingAgentic55.2%2 / 28
OpenHarmony Bench v1.0Coding60.8%1 / 13
OTIS Mock AIME 2024-2025Math99.4% ±0.417 / 200
BabyVisionMultimodal82.0%1 / 11
LiveBench Agentic CodingAgentic64.6%6 / 66
Graduate-Level Google-Proof Q&AKnowledge92.6%15 / 141
RealWorldQAMultimodal88.0%2 / 18
Massive Multi-discipline Multimodal Understanding ProMultimodal82.3%6 / 53
GPQA diamond · Epoch AIKnowledge92.7% ±1.725 / 217
MMMU ProMultimodal88.0% ±0.812 / 92
SWE-bench ProCoding67.7%11 / 83
MMLU ProKnowledge88.6% ±0.318 / 135
MRCRv2Reasoning92.9%2 / 14
SimpleVQAMultimodal75.0%3 / 18
Terminal-Bench 2.1 · BenchLMAgentic86.6%15 / 86
CharXiv Reasoning without toolsMultimodal88.4%3 / 17
WideResearchAgentic81.9%3 / 16
Agents' Last ExamAgenticAbove fitFar above this model's index.52.4%4 / 21
SWE-benchCoding85.6% ±1.617 / 87
LiveBench Instruction FollowingInstruction74.1%13 / 66
AndroidWorldAgentic85.3%2 / 10
ScreenSpot ProMultimodal84.5%4 / 20
LongBench v2Reasoning66.3%5 / 25
GPQA Diamond · BenchLMKnowledge92.6%17 / 84
Mystery Game PuzzlesReasoning38.0% ±4.919 / 80
FrontierMath-Tiers-1-3-v2-PrivateMath74.7% ±2.624 / 90
IOICoding68.9% ±5.212 / 45
NL2RepoCoding55.9%10 / 36
Chess PuzzlesReasoning46.0% ±5.026 / 90
LiveBench ReasoningReasoning88.2%24 / 80
LiveBench Data AnalysisReasoning78.4%24 / 80
Agent Arena task outcomeAgentic4.517 / 54
LiveBench DesignAgentic78.6%17 / 52
FrontierMath-Tier-4-v2-PrivateMath46.3% ±7.924 / 72
Humanity's Last Exam without toolsKnowledge43.6%18 / 53
Terminal-Bench 4.0Agentic34.3% ±3.922 / 61
LiveBench Long AgenticAgentic42.3%19 / 51
Humanity's Last ExamKnowledge43.6%29 / 77
JobBenchAgentic53.4%14 / 36
LiveBench MathematicsMath91.3%26 / 66
Humanity's Last Exam with toolsAgentic56.2%15 / 36
MirrorCodeCoding52.2% ±10.917 / 39
Vibe Code Bench v1.1Coding64.7% ±5.455 / 123
SimpleQA VerifiedKnowledge45.8% ±1.641 / 91
Agent Arena command recoveryAgentic2.823 / 50
Terminal-Bench 2.1 · Vals AIAgentic67.4% ±0.635 / 75
LiveBench LanguageKnowledge79.7%31 / 66
MazeBench: visual spatial reasoning in a 3D open worldAgentic0.0%23 / 48
ProofBench v1.1Math58.0% ±5.026 / 54
Agent Arena steerabilityAgentic−0.828 / 54
ZeroBenchMultimodal24.0%6 / 10
LiveBench ResearchAgentic69.9%33 / 52
Code MigrationCoding24.0% ±4.349 / 77
OSWorld 2.0Agentic19.4%16 / 25
Bug Hunt BenchCoding25.7%17 / 25
Hermes Index: agentic capability and cost per task inside Hermes AgentAgentic36.2%11 / 16
Toolathlon-VerifiedAgentic72.5%16 / 23
ScarfBench: Cross-Framework Application Migration in Enterprise JavaCoding63.7%8 / 11
LiveBench Browser UseAgentic32.0%38 / 49
DeepSWEAgentic56.6%46 / 59
VulcanBench v3Coding81.2%11 / 14
LiveBench CodingCoding72.9%52 / 66
SkillsBenchCoding42.0% ±4.428 / 35
CWE-bench v2Agentic45.4%9 / 11
LiveBench Coding 2.0Coding60.3%46 / 52
Vibe Code Bench 1-100Coding12.8% ±3.223 / 26
FrontierSWE v2Coding15.8%20 / 22
AutomationBenchAgentic27.3%30 / 32
TaxEval v2Finance75.6% ±0.814 / 143
Legal Research BenchLegal47.6% ±3.515 / 90
CorpFin v2Finance65.9% ±0.925 / 131
Tax Agent BenchFinance32.1% ±3.423 / 82
MedScribeHealthcare84.9% ±2.033 / 109
Terminal-Bench-ScienceScience12.9% ±4.014 / 44
Public Benefits Bench v1.1Public benefits67.1% ±1.216 / 50
LegalBenchLegal83.6% ±0.451 / 149
EMBFinance60.1% ±3.435 / 75
MortgageTaxFinance64.0% ±0.949 / 97
ProgramBenchCoding0.0%34 / 65
Finance Agent (v2)Finance50.6% ±0.643 / 79
MedCodeHealthcare40.7% ±2.061 / 106
MysteryMechanismScience23.9% ±2.919 / 30
SAGEEducation81.8% ±2.125 / 29
CyberBench v1.1Other28.6% ±3.348 / 48
MathVision with PythonMultimodal97.7%—
PaperBenchCoding93.0%—
OmniDocBench 1.5Multimodal92.1%—
BabyVision with PythonMultimodal91.3%—
MLVU mean averageMultimodal90.8%—
Video-MME with subtitleMultimodal90.4%—
VideoMMMUMultimodal88.7%—
Multimodal Multi-disciplinary Video UnderstandingMultimodal82.4%—
LVBenchMultimodal81.8%—
MedXpertQA MultimodalMultimodal80.4%—
CC-OCRMultimodal79.6%—
MobileWorldAgentic77.8%—
CoWorkBenchAgentic74.8%—
OCRBench V2Multimodal74.2%—
FrontierSWECoding73.5%—
Hermes Bench: 150 agent tasks graded on the files and state the agent leaves behindAgentic70.5%—
Vision2WebMultimodal69.0%—
WebArena-Verified Browser Agent BenchmarkAgentic66.8%—
PerceptionBench (Internal)Multimodal63.5%—
ZeroBench_main with PythonMultimodal49.0%—
MLS-Bench LiteCoding41.0%—

Sources

BenchLMCC BY-NC 4.0Epoch AICC BYVals AINo licence statedLiveBenchNo licence statedSierra ResearchMITLMArenaCC BY 4.0

History

5 releases
Index67.0 → 65.6
23 Oct 202612 Nov 2026

Questions

How good is Qwen3.8 Max?
Its index is 65.6 out of 100. That is place 40 of 283 ranked models.
What is Qwen3.8 Max best at?
Instruction: 71.4 out of 100, place 2 of 63.
How fast is Qwen3.8 Max?
41 output tokens per second, 52.2 s to the first token.
What is the context window of Qwen3.8 Max?
1M tokens.
Is Qwen3.8 Max open source?
Yes. Its weights are published.
When was Qwen3.8 Max released?
August 2026.

Badge

Qwen3.8 Max on Inference360: place 40 of 283, index 65.6
HTML
<a href="https://inference360.ai/models/qwen3-8-max"><img src="https://inference360.ai/badges/qwen3-8-max.svg" alt="Qwen3.8 Max on Inference360: place 40 of 283, index 65.6" height="20"></a>
Markdown
[![Qwen3.8 Max on Inference360: place 40 of 283, index 65.6](https://inference360.ai/badges/qwen3-8-max.svg)](https://inference360.ai/models/qwen3-8-max)

Compare with

GPT-6 Luna65.5DeepSeek V4 Pro 081365.5GLM-5.365.8Claude Haiku 5.566.3Qwen3.8-27B59.2

More from Alibaba

All models
Qwen3.8-27B59.2Qwen3.7 FlashUnrankedQwen3.8 2.4T A95BUnrankedQwen3.8-27B (Chutes TEE)UnrankedQwen3.8 Max Preview64.2