Alibaba's Qwen3-Max-Thinking outguns Gemini and GPT-5.2

Alibaba just proved that open source darling Qwen can also beat Silicon Valley at closed-source reasoning.

The Alibaba Cloud Qwen team just introduced Qwen3-Max-Thinking, a new proprietary reasoning model designed to go head-to-head with OpenAI and Google. On Humanity’s Last Exam (HLE) — a benchmark packed with 3,000 graduate-level questions designed to be impossible to Google — Qwen3-Max-Thinking paired with web search hit 49.8%, topping both Gemini 3 Pro (45.8%) and GPT-5.2-Thinking (45.5%).

Airbnb CEO Brian Chesky famously praised Qwen’s open weights as a cost-effective OpenAI alternative. With Qwen3-Max-Thinking, Alibaba is no longer just offering cheap open-source substitutes; it wants the enterprise performance crown.

Smarter compute, not just more tokens

Most reasoning models push test-time compute by running brute-force "best-of-N" sampling: generate dozens of candidate answers and rank them.

Alibaba took a different route. Qwen3-Max-Thinking uses an iterative, multi-round reasoning framework backed by a proprietary experience mechanism. Basically, the model learns to spot dead ends before wasting compute exploring them to the end, then funnels those saved cycles directly toward unresolved parts of the problem.

According to Qwen's benchmarks, this stops token consumption from ballooning while boosting accuracy:

- GPQA (PhD-level science) jumped from 90.3 to 92.8. - LiveCodeBench v6 improved from 88.0 to 91.4. - On HMMT Feb 25, it scored 98.0, edging past Gemini 3 Pro (97.5) and beating DeepSeek V3.2 (92.5). - On Arena-Hard v2, it posted a 90.2 score, leaving Claude-Opus-4.5 (76.7) well behind.

Blending deep thought with live tools

Pure reasoning models often struggle when forced to interact with the real world, turning into isolated math savants that cannot browse the web or run scripts.

Qwen3-Max-Thinking bridges this gap by letting its thinking mode invoke tools dynamically without user coaching. It can autonomously juggle live web search, persistent memory, and a Python code interpreter in a single turn.

That means an enterprise agent can search for live quarterly figures, run a regression in Python, and reason through strategic recommendations all in one continuous chain. Grounding logic in external data also directly limits hallucinations.

Why it matters

Western frontier labs spent the last year pitching test-time reasoning as their primary moat. Qwen3-Max-Thinking shows that moat is officially gone.

For developers building complex agentic pipelines, Alibaba's integration of live search and code execution during reasoning is the right architecture. If your workflows require multi-step verification rather than pure math riddles, this is worth benchmarking right now against GPT-5.2 and Gemini 3 Pro.

We think Alibaba's decision to launch a proprietary, aggressive enterprise model is technically justified, but it faces clear commercial hurdles outside Asia. While Airbnb previously leaned on Qwen's open models to cut costs, proprietary Chinese APIs face significant regulatory and security hesitation among Western enterprises, regardless of price.

Still, the broader message is unmistakable: top-tier System 2 reasoning is no longer an American exclusive, and the competition is shifting from raw inference length to tool-integrated efficiency.