Alibaba Releases Qwen4-Max — 2.4T MoE, 71B Active, 1M Context, 62% Cost-Per-Task Drop
Alibaba Releases Qwen4-Max — 2.4T MoE, 71B Active, 1M Context, 62% Cost-Per-Task Drop
Alibaba launched Qwen4-Max on November 16, 2026 — the flagship of its Qwen4 generation. A sparse Mixture-of-Experts (MoE) model with 2.4 trillion total parameters, 71 billion active per token, native vision (text/image/video input), and 1M-token context window. API pricing: $1.50/$4.50 per M tokens (input/output) — a 25% cut from Qwen3.8-Max and ~60% cheaper per task on Alibaba's Work360 benchmark.
Key innovation: Upgraded MoE architecture across attention, residual connections, and embeddings — 25% fewer activated parameters (71B vs 95B) at same 2.4T scale, plus 328B n-gram embedding parameters for local token patterns.
Model Specs: The 2.4T MoE Architecture
| Spec | Qwen4-Max | Qwen3.8-Max | Change |
|---|---|---|---|
| Total parameters | 2.4T | 2.4T | Same backbone |
| Active per token | 71B | ~95B | -25% |
| N-gram embeddings | 328B | — | New component |
| Context window | 1M tokens | 1M | Same |
| Max output | 393K (API) / 262K (thinking) | — | Large |
| Input modalities | Text, images, video | Text only | Native vision |
| Output | Text only | Text only | — |
| MoE design | Upgraded (attention, residual, embeddings) | Standard MoE | Architecture upgrade |
Deployment: Hosted only on QwenCloud API (qwen4-max), Token Plan, Qwen Studio, QwenWork, Qoder, Alibaba Cloud Model Studio (DashScope). No open weights — smaller Qwen4 models promised later.
Pricing: 62% Average Cost-Per-Task Reduction
| Metric | Qwen3.8-Max | Qwen4-Max | Improvement |
|---|---|---|---|
| Input (per 1M) | $2.00 | $1.50 | -25% |
| Output (per 1M) | $6.00 | $4.50 | -25% |
| Implicit cache read | $0.25 | $0.075 | -70% |
| Work360 avg cost/task | $4.43 | $1.67 | -62% |
| Token usage (High effort) | Baseline | -35% | Fewer tokens needed |
| Token usage (XHigh effort) | Baseline | -24% | |
| Terminal-Bench 4.0 token drop | Baseline | -48/49% |
Regional pricing (per 1M tokens):
- Singapore (International): $1.50 in / $4.50 out / $0.075 cache read
- US Virginia, Frankfurt, Tokyo, Hong Kong (Global): $1.413 in / $4.24 out / $0.071 cache read
- China (Beijing): $1.413 in / $4.24 out / $0.071 cache read
API compatibility: OpenAI Chat Completions, OpenAI Responses API, Anthropic Messages API, DashScope — drop-in for existing clients.
Benchmark Performance: Strong Agentic, Uneven Coding
| Benchmark | Qwen4-Max | Competitors | Position |
|---|---|---|---|
| Work360 (721 specialized fields) | Beats Qwen3.8-Max overall, 35% fewer tokens at High effort | — | Strong workplace |
| HLAB-Vals-1.1 (legal agent) | 21.3 | Kimi K3.1: 11.7 | #1 legal agent |
| Code Arena (real-world coding) | 6th | — | Top-tier coding |
| Agents' Last Exam | Strong | — | Agentic workflows |
| Vals Index | Strong | — | |
| JobBench (workplace productivity) | Strong | — | |
| Chartography (visual reasoning) | Strong | — | Native vision |
| FrontierCode 2.0 / hard coding | Behind rivals | Opus 5.5, Fable 5.2, Astra | Gap |
| Vision benchmarks | Behind rivals | GPT-4o, Gemini 4, Fable 5.2 | Gap |
Artificial Analysis Intelligence Index: Not yet listed (Qwen not in AA leaderboard). Estimated ~43 based on Digital Trends — trails Carbon (59), Opus 5.5 (58), Astra (57), Fable 5.2 (~55) but at lower cost.
Competitive Positioning: China's Flagship MoE at Scale
| Model | Provider | Architecture | Active Params | Context | Input/Output | Status |
|---|---|---|---|---|---|---|
| Qwen4-Max | Alibaba | MoE 2.4T | 71B | 1M | $1.50/$4.50 | Live Nov 16 |
| DeepSeek V4.1 Flash | DeepSeek | MoE | — | 128K | ~$0.10/$0.40 | Live (open) |
| GLM 5.3 Flash | Zhipu | — | — | 128K | Competitive | Live |
| Fable 5.2 | Anthropic | Dense | — | 1M | $8/$40 | Live |
| Gemini 4 Carbon | Dense | — | 1M | $2/$10 (intro) | Fairwind | |
| GPT-6 Astra | OpenAI | Dense | — | 1M | ~$10/$50 | Live |
Qwen4-Max's niche: MoE at 2.4T scale with native multimodal, 1M context, API-first distribution across 6 regions (Singapore, US Virginia, Frankfurt, Tokyo, Hong Kong, Beijing). Closed weights — API only.
Cost Advantage: MoE Efficiency + Price Cuts = 62% Task Reduction
The 62% avg cost-per-task drop on Work360 comes from two stacked factors:
- Architectural efficiency — 25% fewer activated params (71B vs 95B) + upgraded MoE = 35% fewer tokens at High effort, 48-49% fewer on Terminal-Bench
- Price cuts — 25% lower list prices + 70% cheaper cache reads = compound savings
vs. Anthropic Opus 5.5 ($4/$20): Qwen4-Max output is ~4.4× cheaper ($4.50 vs $20) vs. Fable 5.2 ($8/$40): Qwen4-Max output is ~8.9× cheaper ($4.50 vs $40) vs. Astra (~$10/$50): Qwen4-Max output is ~11× cheaper
Strategic Implications
- China's MoE leadership at scale — 2.4T MoE with 71B active is the largest deployed sparse model globally
- API-first, closed weights — Alibaba monetizes via hosted API (like OpenAI/Anthropic), not open weights
- Singapore as international hub — QwenCloud International deployment in Singapore for global access
- Regional pricing differentiation — China/Global/International tiers reflect market dynamics
- Open-weight follow-up promised — Smaller Qwen4 models coming open; Max stays closed
- Work360 as proprietary benchmark — 721 specialized fields; industry-task focus vs. academic benchmarks
- Token Plan subscription — Recurring revenue model beyond pay-as-you-go
What Comes Next
- Open-weight Qwen4 models — Smaller variants (7B, 14B, 32B?) promised; will drive adoption
- Work360 independent verification — Third-party evals needed; proprietary benchmark
- Frontier coding/vision gap — Alibaba must close gap to Opus 5.5/Fable 5.2/Gemini 4 on hard coding/vision
- Batch API pricing — Not yet published; critical for high-volume workloads
- Qwen4-Max on neoclouds — OpenRouter, Together, Fireworks, Lambda integration
- Enterprise adoption — Function calling, structured outputs, web search, fine-tuning support
- Regulatory — Singapore deployment navigates export controls; China mainland separate
The Bigger Picture
Qwen4-Max proves MoE at 2.4T scale works in production — not just a research artifact. The 62% cost-per-task reduction (architectural + pricing) shows MoE efficiency compounding with price discipline.
For the global AI race: China's flagship is now MoE-native, multimodal, 1M-context, API-hosted — matching Western labs on architecture but undercutting dramatically on price. The Singapore deployment ensures international access despite export controls.
If Alibaba closes the coding/vision gap with Qwen4.1 or open-weight variants, Qwen becomes the cost leader for agentic workloads — the $1.50/$4.50 price point makes high-volume agent fleets economically viable at unprecedented scale.
Related AIPress Coverage
- DeepSeek's September: V4.1-Flash Beats Its Own Flagship
- Google Gemini 4 Carbon Now Available to Cyber Defenders
- Anthropic Launches Fable 5.2 and Mythos 5.2
- Meta Launches Muse Ember
Sources
- Alibaba Cloud Blog (Nov 16, 2026): "Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery"
- MarkTechPost (Nov 16, 2026): Detailed specs, pricing, benchmark analysis
- VentureBeat (Nov 16, 2026): "Alibaba's Qwen4-Max cuts AI task costs, but benchmarks show uneven challenge"
- Digital Trends (Nov 16, 2026): "Qwen4-Max is China's newest flagship AI model"
- Alibaba Cloud Model Studio Pricing (Nov 17, 2026): Official regional pricing tables
- DataStudios (Nov 16, 2026): API pricing by region, cache types, endpoint compatibility
- LiteLLM (Nov 16, 2026): Qwen4-Max model card, pricing
- APIYI (Nov 17, 2026): Live endpoint verification, 951K recall test
- Digital Trends / The Hindu / Technode / Kingy.ai / ArtificialAnalysis (Nov 2026): Coverage