Alibaba released Qwen3-Max in September 2025 as the flagship of the Qwen3 generation. Where Qwen2.5-Max was primarily a stronger general model, Qwen3-Max introduces a distinct architectural capability: an extended thinking mode that activates chain-of-thought reasoning for complex tasks — analogous to o1-style thinking but available in the same model as standard response mode.
The result is a model that leads LiveCodeBench (86.3), achieves GPQA Diamond 68.1% — the strongest scientific reasoning score in the Chinese lab field at September 2025 — and maintains the Qwen family's pricing and multilingual advantages. For developers evaluating Alibaba models, Qwen3-Max replaces Qwen2.5-Max as the default frontier choice.
What's New
Extended thinking mode: Qwen3-Max supports a thinking parameter that activates longer chain-of-thought reasoning before producing output. In thinking mode, the model works through complex problems step-by-step — math proofs, algorithm design, multi-hop reasoning chains — before delivering a final answer. This is the same paradigm as Claude's extended thinking and OpenAI o1, but integrated into the same model as standard fast response mode.
LiveCodeBench 86.3 — #1 at September 2025: Qwen3-Max leads LiveCodeBench, which evaluates competitive coding problems from LeetCode, Codeforces, and AtCoder. The 86.3 score places it above DeepSeek V4-Flash (approximately 74) and competitive with V4-Pro (93.5). For coding applications where competitive algorithm problems approximate production complexity, this is the metric that matters.
GPQA Diamond 68.1%: The strongest GPQA score among Chinese lab models at September 2025. Graduate-level biology, chemistry, and physics questions require genuine domain knowledge — not pattern matching. Qwen3-Max's 68.1% is above Qwen2.5-Max's 62.9% (+5.2pt) and approaching Claude 3.7 Sonnet-class scientific reasoning.
64K standard context, 128K thinking: Standard mode supports 64K context; extended thinking mode supports up to 128K context with additional thinking budget. The thinking budget parameter controls the token allocation for internal reasoning — higher budget for harder problems.
Qwen3 family composition: Qwen3-Max leads a generation that includes Qwen3-72B (open weight, Apache 2.0), Qwen3-Coder (480B, coding-specialized), and Qwen3-235B-A22B (open weight MoE). The family shares tokenizer and API format, enabling tier-based routing in applications.
Architecture
| Component | Qwen3-Max |
|---|---|
| Architecture | Large-scale MoE (undisclosed parameters) |
| Context (standard) | 64K tokens |
| Context (thinking mode) | 128K tokens |
| Thinking mode | Yes (chain-of-thought with budget control) |
| Languages | 119 languages |
| Input price (standard) | ~¥0.06/K tokens (~$0.083/M) |
| Thinking mode price | Higher (per thinking token) |
| License | Proprietary API |
Extended thinking mechanism: In thinking mode, Qwen3-Max generates an internal reasoning chain before the final answer. The reasoning chain is not returned to the caller by default (reducing output costs), but the model's final answer benefits from the structured reasoning. Thinking budget is specified in tokens — higher budget for more complex problems.
119-language multilingual: A significant expansion from Qwen2.5-Max's 29 languages. Qwen3-Max supports 119 languages via improved multilingual training, though quality is highest for Chinese and English. For applications serving diverse international users, the expanded coverage reduces the need for language-specific model routing.
Qwen3 thinking vs standard mode: Unlike OpenAI's o1 series (which is exclusively a thinking model) or Claude's thinking flag (which applies to all requests), Qwen3-Max allows per-request mode selection. Standard mode returns fast responses at standard pricing; thinking mode activates extended chain-of-thought at higher token cost. This per-request flexibility is useful for applications where most requests are simple (standard mode) but complex requests need deeper reasoning (thinking mode).
Benchmarks
Qwen3-Max at September 2025:
| Benchmark | Qwen3-Max | Qwen2.5-Max | GPT-4o | Claude 3.7S | Notes |
|---|---|---|---|---|---|
| LiveCodeBench | 86.3 | 52.3 | 46.5 | 52.7 | Competitive coding |
| GPQA Diamond | 68.1% | 62.9% | 53.6% | 65.0% | Graduate science |
| Math-500 | 99.1% | 96.9% | 76.6% | 97.3% | Math competition |
| MMLU | 89.7% | 87.9% | 87.2% | 88.3% | General knowledge |
| Arena-Hard | 92.3% | 89.4% | 80.4% | 86.7% | Preference |
LiveCodeBench 86.3 is the defining improvement: +34pt over Qwen2.5-Max (52.3). Thinking mode is the primary driver — the extended chain-of-thought provides structural advantages on algorithm design and competitive programming problems.
GPQA Diamond 68.1% establishes Qwen3-Max as the leading Chinese lab model for scientific reasoning at September 2025. Math-500 at 99.1% (near ceiling) reflects the thinking mode's impact on structured mathematical proofs.
Pricing in Context
| Model | Input $/MTok | Thinking Mode | LiveCodeBench |
|---|---|---|---|
| Qwen3-Max | ~$0.083 | Yes | 86.3 |
| Qwen2.5-Max | ~$0.055 | No | 52.3 |
| Claude 3.7 Sonnet | $3.00 | Yes | 52.7 |
| GPT-4o | $5.00 | No | 46.5 |
| DeepSeek V4-Pro | $0.435 | Yes (xhigh) | 93.5 |
Qwen3-Max at ~$0.083/M with thinking mode is the most cost-effective option for GPQA Diamond-class scientific reasoning at September 2025. DeepSeek V4-Pro leads on LiveCodeBench (93.5 vs 86.3) at $0.435/M — a meaningful gap for competitive coding specifically, but Qwen3-Max's Chinese language strength and Alibaba ecosystem integration are the counterweights.
What It Means for Developers
Per-request thinking budget: Unlike separate thinking models, Qwen3-Max's per-request mode selection means a single API call can specify thinking level based on request complexity. Routing logic: use standard mode for factual questions and short completions, thinking mode for math, code design, and multi-step reasoning. This reduces average cost while maintaining quality on hard cases.
GPQA-class scientific applications: 68.1% GPQA Diamond positions Qwen3-Max as the Chinese lab model of choice for research assistants, scientific literature synthesis, and domain-expert AI tools in biology, chemistry, and physics. At $0.083/M with thinking mode, the economics support research AI applications that would be expensive at Claude or GPT-4o pricing.
Qwen3 family routing: Qwen3-Max (hard reasoning), Qwen3-Coder (agentic coding), Qwen3-72B (open-weight self-hosted), and Qwen3-235B-A22B (large open MoE) share the same API format. A single application can route by task type — max for analysis, coder for code edits, 72B for high-volume classification — with no integration differences.
119-language coverage for global apps: The expanded multilingual support reduces language-routing complexity for applications serving global users. Standard and thinking modes both support the full 119-language range.
Availability
- API: Alibaba Cloud Model Studio —
qwen-max(Qwen3-Max) - Web: Tongyi Qianwen — powered by Qwen3-Max
- Qwen3-72B open weights: HuggingFace — Apache 2.0
- Pricing: ~¥0.06/K tokens standard, higher for thinking tokens
Bottom Line
Qwen3-Max is a material upgrade over Qwen2.5-Max — LiveCodeBench +34pt, GPQA Diamond +5.2pt, Math-500 near-ceiling, per-request thinking mode. At ~$0.083/M with thinking mode, it's the strongest Chinese lab option for scientific reasoning and complex problem-solving in September 2025. For developers on Alibaba Cloud, Qwen3-Max replaces Qwen2.5-Max as the default frontier choice — the same ecosystem, same API format, significantly stronger on the benchmarks that measure hard reasoning.
Resources
- Alibaba Cloud Model Studio — Qwen3-Max API
- Qwen HuggingFace — Qwen3 open models
- Qwen Blog — Qwen3 release post
- Tongyi Qianwen — web interface
Qwen3-Max is available via Alibaba Cloud Model Studio at dashscope.aliyuncs.com.
