Unspecified Reasoning (With Tools) · 19 ranked
Qwen4-Max
Alibaba's closed-weight flagship multimodal reasoning model (Nov 2026): 2.4T backbone / 71B active MoE + 328B N-gram embeddings, 1M context, 393K output, $1.50/$4.50 per 1M tokens.
At a glance
- Context window
- 1M tokens
- Max output
- 384K tokens
- Text input / output · per 1M tokens
- $1.50 / $4.50
- Knowledge cutoff
- —
- BioMysteryBench
- #1 of 28 models on the BioMysteryBench leaderboard
- Agents' Last Exam v1.1
- #1 of 17 models on the Agents' Last Exam v1.1 leaderboard
Benchmark snapshot
Best result in 6 of 7 categories
AcademicHLE
51.7Max (No Tools) · #42 of 151
CodingDeepSWE
75.7Max (With Tools) · #5 of 73
Computer UseAgents' Last Exam v1.1
46.7Max (With Tools) · pass-rate · #1 of 17
General EvaluationAA Intelligence Index v5.0.1
42.9Max (task-specific tools) · #18 of 62
MultimodalChartography
54.2Max (With Tools) · #23 of 29
ProfessionalAA-Briefcase
1686EloMax (With Tools) · #6 of 20
Benchmark details
19 results · Best published score per benchmark, grouped by category
| Benchmark · setting | Score | Rank |
|---|---|---|
| Academic | ||
| HLE Max (No Tools) · Knowledge Exams | 51.7 | 42 / 151 |
| Coding | ||
| DeepSWE Max (With Tools) · Repository Engineering | 75.7 | 5 / 73 |
| ProgramBench Max (With Tools) · programbench-almost-resolved · Code Generation & Editing | 31.5 | 7 / 23 |
| Terminal-Bench 4.0 Unspecified Reasoning (With Tools) · vals-mini-swe · limited comparability · Agentic Development Source: Vals AI, Terminal-Bench 4.0 | 48.5 | 20 / 90 |
| Vibe Code Bench v1.1 Max (With Tools) · Code Generation & Editing Source: Vals AI — Vibe Code Bench v1.1. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 86.7 | 16 / 121 |
| Computer Use | ||
| Agents' Last Exam v1.1 Max (With Tools) · pass-rate · Tool Orchestration | 46.7 | 1 / 17 |
| APEX-Agents Max (With Tools) · Tool Orchestration | 72.1 | 5 / 61 |
| AutomationBench-AA Max (With Tools) · Tool Orchestration | 72.5 | 9 / 18 |
| General Evaluation | ||
| AA Intelligence Index v5.0.1 Max (task-specific tools) · Capability Indices | 42.9 | 18 / 62 |
| Agent Arena Max (With Tools + Internet) · Preference Arenas | 7.5 | 12 / 35 |
| Agent Arena Index Extra-High (Tools + Internet) · Preference Arenas Source: Arena Agent Arena Index | 1189 | 15 / 36 |
| Code Arena WebDev Thinking (With Tools) · Preference Arenas | 1750Elo | 1 / 4 |
| Vals Index v2.2 Max (With Tools) · Capability Indices Source: Vals AI — Vals Index v2.2. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 63.2 | 13 / 36 |
| Multimodal | ||
| Chartography Max (With Tools) · Documents & Charts | 54.2 | 23 / 29 |
| Professional | ||
| AA-Briefcase Max (With Tools) · Office & Business | 1686Elo | 6 / 20 |
| JobBench Max (With Tools) · Cross-industry Work | 68.5 | Unranked |
| Legal Research Bench Max (With Tools) · Legal Source: Vals AI — Legal Research Bench. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 41.4 | 32 / 91 |
| Tax Agent Bench Max (With Tools) · Finance | 40.4 | 29 / 74 |
| Science & Health | ||
| BioMysteryBench Max (With Tools) · Biology & Genomics | 85.0 | 1 / 28 |
Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.
Back to benchmark snapshot ↑API price vs benchmark score
Methodology & notes (1 unpriced omitted)
Uses listed uncached text API rates at a 4:1 input/output mix against one published protocol and setting (programbench-almost-resolved). API rates are separate from evaluation token costs and do not measure cost per solved task. Selected point basis: International pricing. Up to 1,000,000 input tokens QwenCloud International (Singapore)
API pricing
Listed text rates, USD per million tokens · effective 2026-11-16
- Input
- $1.50
- Output
- $4.50
- Cached input
- $0.075
Blended price4:1 input:output · 90% of input tokens cached · same basis as the price ranking
$1.074/ 1M tokens
Specifications
- Provider
- Alibaba
- Model type
- Reasoning model
- Released
- November 16, 2026
- Total parameters
- 2.4T
- Active parameters
- 71B
- Architecture
- Mixture of experts
- Context window
- 1M tokens
- Max output
- 384K tokens
- Input
- Text, Image, Video
- Output
- Text
- Weights
- Closed / API only
- Commercial use
- Proprietary
Capabilities
Model analysis
Release Overview
Qwen4-Max is Alibaba's closed-weight flagship multimodal reasoning model in the Qwen4 generation, published on the official Qwen blog on November 15, 2026 and rolled out for general API availability across QwenCloud and Qianwen AI (following a five-day qwen4-max-preview window on Token Plan starting November 11). It keeps a 2.4T-parameter sparse MoE backbone while reducing active parameters from 95B in Qwen3.8-Max to 71B, and adds 328B N-gram Embedding parameters. Across 16 benchmarks reported in the launch blog, Qwen4-Max leads on DeepSWE v1.1 (75.7), Agents' Last Exam v1.1 (46.7), HLAB-Vals-1.1 (21.3), and BioMysteryBench (85.0), while trailing Opus 5.5, GPT-6.1 Sol, or Kimi K3.1 on Terminal-Bench 4.0 (48.5), ProgramBench (31.5), and Humanity's Last Exam (51.7).
71B Active MoE with 328B N-gram Embeddings and 393K Output
Per the official blog (qwen.ai/blog?id=qwen4-max), Qwen4-Max pairs three Gated DeltaNet layers for every Qwen Sparse Attention (QSA) layer with micro-block indexing, expands the residual stream into four parallel Gated Residual branches, and adds a 328B N-gram Embedding lookup table at layer 2. On QwenCloud (qwencloud.com), it accepts native text, image, and video inputs with a 1M-token context window (991K max input standard, 983K in thinking mode) and up to 393,216 output tokens shared with reasoning (up to 262K CoT), configurable via enable_thinking and effort levels.
Strong Gains in Legal, Bio, and Agentic Work; Gaps in Terminal and Black-Box Coding
Compared against Qwen3.8-Max-0902, Qwen4-Max improves across all 16 reported benchmarks, though its standing against frontier peers splits sharply by domain:
- Leading: DeepSWE v1.1 at 75.7 (vs. Kimi K3.1 75.5, GPT-6.1 Sol 75.2, Opus 5.5 74.2); Agents' Last Exam v1.1 at 46.7 (vs. Opus 5.5 / Kimi K3.1 43.4); HLAB-Vals-1.1 at 21.3 (vs. Kimi K3.1 11.7, Opus 5.5 9.2); BioMysteryBench at 85.0; and Alibaba's Work360 at 52.69 (up from 47.74).
- Competitive: APEX-Agents v1.1 at 72.1 (Opus 5.5: 73.5), Automationbench-AA at 72.5 (Kimi K3.1: 73.3), AA-Briefcase v1.1 at 1686 Elo (Opus 5.5: 1801), JobBench at 68.5 (Opus 5.5: 75.4), CodeArena at 1750 Elo (Opus 5.5: 1816), and HealthBench Professional at 60.0 (GPT-6 Astra: 66.0).
- Trailing: Terminal-Bench 4.0 at 48.5 (Kimi K3.1: 67.2, Opus 5.5: 59.6), ProgramBench at 31.5 (Opus 5.5: 65.0), AgentArena net improvement at 7.5 (Opus 5.5: 18.2), Humanity's Last Exam (no tools) at 51.7 (Opus 5.5: 59.8), Chartography at 54.2 (GPT-6 Astra: 71.1), and Video-Index at 68.1 (GPT-6 Astra: 76.7).
On secure coding (Last Secure Code Bench), functional+security pass rate rose from 22.0% to 31.8%. The release also introduces Live Delivery for long-running agents and open-sources Qwen-Live-Delivery-Plugin.
Lower API Pricing and 62% Lower Per-Task Cost
On QwenCloud International (Singapore), Qwen4-Max lists at $1.50 per million input tokens and $4.50 per million output tokens (25% below Qwen3.8-Max), with both implicit and explicit cache-hit reads at $0.075/M (70% lower), 5-minute explicit cache creation at $1.875/M, and 1-hour explicit cache creation at $3.00/M. In mainland China (Qianwen AI), standard rates are ¥10 / ¥30 per million tokens, ¥0.5/M for cache hits, ¥12.5/¥20 per million for 5-min/1-hr cache writes, and ¥5/¥15 for Batch File. Combined with ~24%–35% lower token usage on Work360 and ~48%–49% on Terminal-Bench 4.0, Alibaba reports Work360 average per-task cost dropped 62% from $4.43 to $1.67.
Closed Weights and Selection Boundaries
Qwen4-Max is closed-weight at launch (OpenAI- and Anthropic-compatible APIs); Alibaba stated smaller Qwen4 models will be open-sourced later. Teams needing downloadable weights can still use Qwen3.8-Max (Qwen3.8-2.4T-A95B), while terminal-heavy and black-box replication workloads still favor Kimi K3.1 or Opus 5.5. For cache-heavy multimodal agents, legal/bio workflows, and 393K-output tasks, Qwen4-Max offers strong price-performance.
Sources
Last reviewed 2026-11-16
Questions
- What are Qwen4-Max's parameter count and context limits?
- Qwen4-Max uses a 2.4T-parameter sparse MoE architecture with 71B active parameters per token. It supports a 1M-token context window (991K max input standard, 983K in thinking mode) and up to 393,216 output tokens shared with reasoning.
- What is the API pricing for Qwen4-Max?
- On QwenCloud International, Qwen4-Max costs $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 per million implicit cache-hit input tokens. In mainland China (Qianwen AI), standard rates are ¥10 input / ¥30 output / ¥0.5 cache hit per million tokens.
- Are Qwen4-Max weights open source?
- No. At its November 16, 2026 launch, Qwen4-Max is available only via closed-weight API and Alibaba's platforms. Alibaba stated that smaller Qwen4 models will be open-sourced later.
Qwen4-Max vs competitors
Comparable results only · record is Qwen4-Max's wins–losses
| Benchmark | Qwen4-MaxThis model | vs Kimi K3.1 ProRecord 11–4–1 | vs DeepSeek-V4.1-Flash-1017Record 9–6 | vs Claude Opus 5.5Record 3–12 | vs Gemini 4 ArgonRecord 2–12 |
|---|---|---|---|---|---|
| Vibe Code Bench v1.1 | 86.7 | 84.9 | 87.7 | 90.3 | 91.9 |
| Legal Research Bench | 41.4 | 37.5 | 45.2 | 50.5 | 54.8 |
| Tax Agent Bench | 40.4 | 24.9 | 44.4 | 70.5 | 44.6 |
| DeepSWE | 75.7 | 75.5 | 77.0 | 74.2 | 77.9 |
| AA Intelligence Index v5.0.1 | 42.9 | 46.5 | 32.6 | 59.1 | 50.4 |
| APEX-Agents | 72.1 | 67.2 | 63.4 | 73.5 | 82.4 |
| Agent Arena Index | 1189 | 1145 | 1109 | 1336 | 1229 |
| Vals Index v2.2 | 63.2 | 62.5 | 61.3 | 70.2 | 69.2 |
| Agent Arena | 7.5 | 4.8 | 2 | 19.5 | 10.7 |
| Chartography | 54.2 | 90.2 | 83.0 | 89.0 | 71.6 |
Lineage
Predecessors and successors · oldest to newest
- Qwen3.5 Max PreviewFeb 1, 2026
- Qwen3.6-Max-PreviewApr 18, 2026
- Qwen3.7 MaxMay 20, 2026
- Qwen3.8-Max-PreviewJul 19, 2026
- Qwen3.8-MaxAug 3, 2026
Generational jump vs Qwen3.8-Max
1 like-for-like benchmark with the same protocol and setting.
Price & benchmark trends
Across curated predecessors and successors, ordered by release date. Points use listed rates and published results, not a history of price changes within one model. Related branches may appear; adjacent points do not imply direct succession. Undated models appear last.
API price across generations
Input and output share one linear USD axis · uncached text rates · lower is cheaper. Retired models may retain their last recorded price. Rate tiers and context limits can differ; inspect each model's pricing conditions.
Benchmark performance across generations
Resolution rate (%) · Higher is better. Same recorded metric, protocol and setting; review exclusions are respected. Score gaps do not establish statistical significance.
Inspect Terminal-Bench 4.0 →Orange marks this model. Hover, tap or focus the chart to see exact values. Missing values leave gaps. Horizontal spacing shows release order, not elapsed time.
| Model · released | Input / 1M | Output / 1M | Price effective | Terminal-Bench 4.0 |
|---|---|---|---|---|
| 1Qwen3.5 Max PreviewFeb 1, 2026 | — | — | Not recorded | — |
| 2Qwen3.6-Max-PreviewApr 18, 2026 | $1.30 | $7.80 | 2026-04-18 | — |
| 3Qwen3.7 MaxMay 20, 2026 | $2.50 | $7.50 | 2026-11-13 | — |
| 4Qwen3.8-Max-PreviewJul 19, 2026 | — | — | Not recorded | — |
| 5Qwen3.8-MaxAug 3, 2026 | $2.00 | $6.00 | 2026-11-13 | 34.3 ↗ |
| 6Qwen4-MaxNov 16, 2026 · This model | $1.50 | $4.50 | 2026-11-16 | 48.5 ↗ |
Primary sources
- Qwen4-Max technical reportqwen.ai
- Qwen4-Max online demoqianwenai.com
Community discussion
Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.
Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.
No community comments yet.
