Atlas

Models

Alibaba

Qwen4-Max

164.4
Capability · 95% interval 162.7–166.3
Rank 20 of 899 (interval 15–25)
Relative cost 1.1× · $0.098 per short-answer task · $0.84 per agentic task

Qwen4-Max is Alibaba's hosted flagship Qwen4 model, released on November 16, 2026 as qwen4-max on Alibaba Cloud Model Studio, which lists no dated snapshot. Qwen reports a 2.4 trillion-parameter backbone with 71 billion active parameters, plus 328 billion N-gram embedding parameters. The API accepts text, images and video, returns text, and supports a 1,000,000-token context; thinking is on by default, and reasoning_effort takes low, high or xhigh, with high the default. The weights are closed, and Qwen says smaller Qwen4 models will be open-sourced later. A separate qwen4-max-preview endpoint ran from November 11 on Alibaba's Token Plan.

Scored on 167 benchmarks · 166 total scores · 166 independently reported

Score providers

Organizations reporting headline scores for 167 benchmarks

Qwen4-Max benchmark scores

In fitin the capability model (18 of 167)
Legal · 69 benchmarks · first place on 4
HLAB-Vals-1.1 - Environmental ESGx-high
1 / 40———
HLAB-Vals-1.1 - Taxx-high
1 / 40———
HLAB-Vals-1.1 - Criteria Pass Rate - Arbitration International Dispute Resolutionx-high
1 / 40———
HLAB-Vals-1.1 - Criteria Pass Rate - Data Privacy Cybersecurityx-high
1 / 40———
HLAB-Vals-1.1 - Antitrust Competitionx-high
2 / 40———
Finance & Accounting · 35 benchmarks · first place on 1
Finance Agent v2.1 - General Qualitative Analysisx-high
1 / 40———
Finance Agent v2.1 - All Passx-high
2 / 40———
Finance Agent v2.1 - Adjustmentsx-high
2 / 40———
Excel Modeling Benchmark v1.1 Miscx-high
3 / 44———
Finance Agent v2.1 - Precedentsx-high
4 / 40———
Code · 16 benchmarks
Terminal-Bench 4.0 — MLx-high
3 / 59———
Terminal-Bench 4.0 — Operationsx-high
3 / 59———
Arena Image-to-WebDevIn fitx-high
1674Arena
5 / 49~1 mo—2026-11-15
LMArena Code Arena WebDevIn fitx-high
1750Arena
6 / 117~2 mo—2026-11-15
Arena Code WebDev — Reactx-high
1785Arena
6 / 104——2026-11-15
Agents · 12 benchmarks
EnterpriseOps-Gym-AAIn fitx-high
4 / 52~1.6 yearsNo trend—
AA-Briefcase v1.1 EloIn fitx-high
5 / 169~3 mo——
AutomationBench-AA Operations, HR, Marketing, Sales and Supportx-high
5 / 26———
AutomationBench-AA Support and Operationsx-high
6 / 26———
AutomationBench-AA 1.0.6In fitx-high
7 / 163~2 mo~May 2027—
Indexes · 12 benchmarks
Vals Finance Indexx-high
6 / 40———
Artificial Analysis Finance & Accounting Indexx-high
7 / 26———
Artificial Analysis Legal Indexx-high
7 / 26———
Artificial Analysis Economics Indexx-high
7 / 26———
Artificial Analysis Strategy & Ops Indexx-high
8 / 26———
Professional Work · 9 benchmarks · first place on 1
APEX-Agents-AA (avg@3)x-high
1 / 34———
MLCR-AA Completeness (October 2026 methodology)x-high
6 / 114———
GDPval-AA v2.1x-high
6 / 209———
GDP.pdf (Artificial Analysis, All-pass)In fitx-high
14 / 162~2 mo~Jan 2029—
GDP.pdf (Artificial Analysis, Mean Criteria)x-high
14 / 162———
General QA · 7 benchmarks
Humanity's Last Exam Diamond (Knowledge Partition)x-high
10 / 101———
AA-Omniscience Law Accuracyx-high
10 / 26———
AA-Omniscience Business Accuracyx-high
11 / 26———
AA-Omniscience Health Accuracyx-high
11 / 26———
Humanity's Last Exam Diamond (Reasoning Partition)x-high
14 / 101———
Safety · 4 benchmarks
AA-Omniscience Law Non-Hallucinationx-high
6 / 26———
AA-Omniscience Business Non-Hallucinationx-high
7 / 26———
AA-Omniscience Health Non-Hallucinationx-high
14 / 26———
AA-Omniscience v1.1 Hallucination Ratex-high
22 / 76———
Science · 2 benchmarks
Terminal-Bench-Science 0.1In fitx-high
17 / 78~1 mo~Mar 2027—
CritPt v2 (agentic, avg@5)In fitx-high
22 / 77~2 mo~Mar 2027—
Health & Medicine · 1 benchmark
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)In fitx-high
34 / 114~4 mo~Mar 2029—

Related models

Loading Atlas data…