The first model of the Qwen4 generation, released November 16, 2026 as a hosted flagship. It is built on the new Qwen4 architecture that Qwen3.8-Flash-Next previewed in open weights. Specifications:

  • Size: a 2.4T-parameter sparse MoE backbone with 71B active per token, plus 328B parameters of N-gram embeddings.
  • Modality and context: native vision input and a 1M-token context, with up to 393K output tokens.
  • Price: $1.50 / $4.50 per million input/output tokens, against Qwen3.8-Max's $2 / $6. Cached input costs $0.075.

No weights are released for the Max model. Alibaba says smaller Qwen4 models from the same lineage are "coming to open source".

Architecture. The blog lists three upgrades over Qwen3.8-Max:

  • Attention: each Qwen Sparse Attention (QSA) layer is preceded by three Gated DeltaNet layers that compress history into fixed-size states. QSA's lightweight indexer then selects relevant context at micro-block granularity.
  • Gated Residual: four parallel residual branches with dynamic read and write gates.
  • N-gram embedding: a context-based lookup of local token patterns at the second layer.

Qwen reports a pretraining run with no loss spikes. Qwen4-Max-Base has 5–13% lower bits per byte than Qwen3.8-Max-Base in every in-house domain (knowledge, code, QA, STEM) while activating 25% fewer parameters (71B against 95B). Training tokens and compute are not disclosed, and there is no technical report.

Artificial Analysis scores it 43 on Intelligence Index v5.0.1, at $3.27 per index task. That is up from Qwen3.8-Max's 34, and second among Chinese models to Moonshot's Kimi K3.1 Pro (46).

Self-reported results, compared with Qwen3.8-Max-0902, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1:

  • Best in the chart: DeepSWE v1.1 (75.7), Agents' Last Exam v1.1 (46.7, against Opus 5.5's 43.4), HLAB-Vals 1.1 legal work (21.3), BioMysteryBench (85.0) and HealthBench Professional (60.0).
  • Behind the leaders: Terminal-Bench 4.0 (48.5, against 59.6 for Opus 5.5 and 67.2 for Kimi K3.1), ProgramBench (31.5 against 65.0), AA-Briefcase v1.1 (Elo 1686 against 1801), JobBench (68.5 against 75.4), Humanity's Last Exam (51.7 against 59.8) and chart understanding (54.2 against 71.1).

It also uses fewer tokens than Qwen3.8-Max: 24–35% fewer on Alibaba's new Work360 professional-task benchmark (52.69, up from 47.74) and about 48% fewer on Terminal-Bench 4.0. Work360 is built on a taxonomy of 26 domains, 182 subdomains and 721 fields. The average Work360 cost per task fell 62%, from $4.43 to $1.67.

Alibaba presents it as a "24/7 AI teammate". Its showcases include:

  • Blender modeling, and games built in Unreal, Unity and Godot.
  • Automotive production-line planning and machining. Programming one workpiece went from about 475 minutes to 24.
  • Legal due diligence and financial research with 90 MCP tools.
  • Closed-loop science: a 48-hour autonomous run on Deep Principle's platform designed, synthesized and validated a new data-center coolant molecule. On DP Technology's platform, it designed an NLRP3 inhibitor candidate.
  • Optimizing Alibaba Cloud's own GPU and CPU scheduling.

"Live Delivery" ships long-running agent work in staged releases the user can redirect, and its plugin is open source. The safety section reports secure-code pass rates rising from 22.0% to 31.8% on Last Secure Code Bench, and finds no large-scale cyber or CBRN uplift. Qwen4-Max is available through the QwenCloud API and Token Plan, QwenWork, Qoder and QwenStudio.

Model Details

Architecture MOE
Parameters 2.4T
Active params 71B
Context window 1,000,000
AA Intelligence 43
License Proprietary (API)

Benchmark Scores

Benchmark Score Mode
DeepSWE v1.1 75.7 —
Terminal-Bench 4.0 48.5 —
Agents' Last Exam v1.1 46.7 —
AA-Briefcase v1.1 1686 Elo —
AutomationBench-AA 72.5 —
APEX-Agents v1.1 72.1 —
Humanity's Last Exam 51.7 —
Work360 52.69 —
frontiermoereasoningagenticcodingmultimodallong-contextproprietary

Related