Inference
EN
Alibaba

Qwen3.5 397B

Qwen3.5 397B is a standard model from Alibaba. 46 benchmarks count toward its index, in 8 categories.

  • Open weights
  • Standard
  • Input: text
  • Released 2026-02
RankedEnough results for an index.
IndexAll categories, out of 100.53.8±2.4Median 51.0
CoverageIndex weight that has results.100%
SpeedOutput tokens per second.59/sMedian 67/s
First tokenSeconds to the first output token.56.6 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$0.60Median $0.60
Output / 1MUS dollars per 1M output tokens.$3.60Median $2.55
ContextMaximum tokens in one request.128KMedian 262K
EloLMArena rating and rank.—Median 1422

Run locally

403.4B parameters
16-bit
848.1 GB
8-bit
451.0 GB
4-bit
257.8 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.13 counted53.3110 / 218
CodingCode writing and repair.5 counted53.3113 / 209
ReasoningLogic problems and puzzles.3 counted51.9102 / 227
MultimodalTasks with images and text.11 counted55.652 / 84
KnowledgeFacts and expert knowledge.6 counted53.8104 / 268
MultilingualTasks in many languages.2 counted55.63 / 9
InstructionTasks with strict prompt rules.2 counted55.623 / 63
MathMath problems.4 counted53.672 / 168

Results

46 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
Artificial Analysis IFBenchInstruction78.8%6 / 146
Massive Multitask Language Understanding ProfessionalKnowledge87.8%5 / 75
τ²-Bench Tool-Agent-User EvaluationAgentic95.6%18 / 149
Massive Multi-discipline Multimodal UnderstandingMultimodal85.0%4 / 28
MMLU-ReduxKnowledge94.9%3 / 19
RefCOCO averageMultimodal92.3%2 / 10
V*Multimodal95.8%3 / 14
Artificial Analysis GPQA DiamondKnowledge89.3%45 / 199
Instruction-Following EvalInstruction92.6%16 / 70
MMLU-ProXMultilingual84.7%4 / 17
C-EvalKnowledge93.0%3 / 12
τ³-Bench Tool-Agent-User EvaluationAgentic68.4%8 / 32
Graduate-Level Google-Proof Q&AKnowledge88.4%36 / 141
LongBench v2Reasoning63.2%7 / 25
CountBenchMultimodal97.2%3 / 10
Artificial Analysis MMMU-ProMultimodal77.3%40 / 113
Artificial Analysis Long Context ReasoningReasoning77.3%77 / 217
Massive Multi-discipline Multimodal Understanding ProMultimodal79.0%19 / 53
Software Engineering Benchmark VerifiedCoding76.2%42 / 114
Artificial Analysis Humanity's Last ExamKnowledge29.0%85 / 222
VITA-BenchAgentic43.7%5 / 13
Harvard-MIT Mathematics Tournament February 2025Math94.8%6 / 15
Harvard-MIT Mathematics Tournament February 2026Math87.9%11 / 27
SuperGPQA: Scaling LLM Evaluation Across 285 Graduate DisciplinesKnowledge70.4%9 / 22
MathVisionMultimodal88.6%10 / 24
INCLUDEMultilingual85.6%5 / 12
LiveCodeBench v6Coding83.6%25 / 59
Instruction Following BenchmarkInstruction76.5%25 / 57
RealWorldQAMultimodal83.9%8 / 18
Harvard-MIT Mathematics Tournament November 2025Math92.7%6 / 12
ERQAMultimodal67.5%9 / 18
Critical Physics TasksReasoning1.7%109 / 214
Gert Labs Composite Game BenchmarkAgentic46.8%27 / 52
SimpleVQAMultimodal67.1%10 / 18
AIME 2026Math93.3%21 / 34
WideResearchAgentic74.0%10 / 16
CharXiv ReasoningMultimodal80.8%31 / 49
Terminal-Bench 2.0 · BenchLMAgentic52.5%26 / 41
Artificial Analysis SciCodeCoding44.8%74 / 114
Humanity's Last ExamKnowledge28.7%52 / 77
BrowseCompAgentic69.0%43 / 63
APEX-Agents-AAAgentic15.3%21 / 30
ScreenSpot ProMultimodal65.6%15 / 20
Claw-EvalAgentic56.8%34 / 42
SWE-bench ProCoding50.9%69 / 83
OSWorld-VerifiedAgentic62.2%31 / 37
ToolathlonAgentic36.3%19 / 22
MCP AtlasAgentic46.1%47 / 52
NL2RepoCoding32.2%33 / 36
QwenClawBenchAgentic51.8%10 / 10
AI2D test splitMultimodal93.9%—
OmniDocBench 1.5Multimodal90.8%—
MAXIFEMultilingual88.2%—
We-MathMultimodal87.9%—
Video-MME with subtitleMultimodal87.5%—
MLVU mean averageMultimodal86.7%—
DynaMathMultimodal86.3%—
VideoMMMUMultimodal84.7%—
Video-MME without subtitleMultimodal83.7%—
CC-OCRMultimodal82.0%—
MMAnswerBenchMath80.9%—
MCP-TasksAgentic74.2%—
PolyMathMultilingual73.3%—
AI-NeedleReasoning68.7%—
MMLongBench-DocMultimodal61.5%—
NOVA-63Multilingual59.1%—
ODINW13Multimodal47.0%—
DeepPlanningAgentic37.6%—
ResearchClawBenchAgentic14.2%—

Sources

BenchLMCC BY-NC 4.0

History

5 releases
Index53.0 → 53.8
23 Oct 202612 Nov 2026
Output / 1M$3.60 → $3.60
23 Oct 202612 Nov 2026

Questions

How much does Qwen3.5 397B cost?
$0.60 per 1M input tokens, $3.60 per 1M output tokens.
How good is Qwen3.5 397B?
Its index is 53.8 out of 100. That is place 109 of 283 ranked models.
What is Qwen3.5 397B best at?
Multilingual: 55.6 out of 100, place 3 of 9.
How fast is Qwen3.5 397B?
59 output tokens per second, 56.6 s to the first token.
What is the context window of Qwen3.5 397B?
128K tokens.
Is Qwen3.5 397B open source?
Yes. Its weights are published.
When was Qwen3.5 397B released?
February 2026.

Badge

Qwen3.5 397B on Inference360: place 109 of 283, index 53.8
HTML
<a href="https://inference360.ai/models/qwen3-5-397b"><img src="https://inference360.ai/badges/qwen3-5-397b.svg" alt="Qwen3.5 397B on Inference360: place 109 of 283, index 53.8" height="20"></a>
Markdown
[![Qwen3.5 397B on Inference360: place 109 of 283, index 53.8](https://inference360.ai/badges/qwen3-5-397b.svg)](https://inference360.ai/models/qwen3-5-397b)

Compare with

GPT-5.1-Codex-Max53.8GPT-5.4 mini54.5Qwen3.6 Plus54.6Qwen3.5 397B (Reasoning)55.3Qwen3.5-122B-A10B51.9

Lower price

dots3-note Preview63.4 · FreeApodex 1.1 Mini55.7 · FreeNemotron 3 Ultra50.8 · Free

More from Alibaba

All models
Qwen3.5 397B (Reasoning)55.3Qwen3.5 397B A17BUnrankedQwen3.5 Plus 2026-02-15UnrankedQwen3 Max ThinkingUnrankedQwen3 Coder NextUnranked