Alibaba Cloud's Qwen team has unveiled Qwen3-Max-Thinking, a closed-source language model focused on complex reasoning that outperforms leading models such as GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3 Pro across a range of benchmarks. The model combines a massive parameter count with advanced machine learning techniques and offers developers an affordable alternative to Western solutions.
Architecture based on test-time scaling
The key innovation of Qwen3-Max-Thinking is a technique called "test-time scaling". Unlike standard methods, where the model generates tokens linearly, Qwen3 uses a so-called "heavy mode," which allows the model to trade compute for intelligence. Instead of the naive "best-of-N sampling" approach, where the model generates, for example, 100 responses and selects the best one, Qwen3-Max-Thinking uses a multi-round cumulative experience strategy. This approach mimics human problem-solving. The model uses a proprietary "take-experience" mechanism to distill insights from previous reasoning steps, allowing it to identify dead ends and redirect compute toward unresolved ambiguities.
The results are tangible. On the GPQA benchmark (PhD-level science questions), the score improved from 90.3 to 92.8. On LiveCodeBench v6, performance increased from 88.0 to 91.4, all without a proportional increase in token costs.
Adaptive tool use for practical applications
Qwen3-Max-Thinking bridges the gap between pure "thinking" and practical use by integrating three tools as first-class capabilities: web search, memory, and a code interpreter. The model autonomously selects the right tool for a given task without manual input from the user. In "Thinking Mode," the model can use these tools simultaneously. This capability is critical for enterprise applications, where the model may need to verify a fact (search), calculate a projection (code interpreter), and then reason about the strategic implications (thinking) - all in a single step.
The Qwen team notes empirically that this combination "effectively mitigates hallucinations" because the model can ground its reasoning in verifiable external data instead of relying solely on its training weights.
Benchmark results
Across 19 public benchmarks, Qwen3-Max-Thinking performed on par with or better than GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3 Pro. In knowledge tests, the model achieved 85.7 on MMLU-Pro, 92.8 on MMLU-Redux, and 93.7 on C-Eval, where it led the group in Chinese-language evaluation. For complex reasoning, it recorded 87.4 on GPQA, 98.0 on HMMT Feb 25, 94.7 on HMMT Nov 25, and 83.9 on IMOAnswerBench, placing it among the top tier of current mathematical and scientific models.
The most significant signal for developers comes from the "Humanity's Last Exam" (HLE) benchmark, which measures performance on 3,000 graduate-level questions spanning mathematics, science, computer science, the humanities, and engineering. Qwen3-Max-Thinking equipped with web search tools achieved a score of 49.8, beating both Gemini 3 Pro (45.8) and GPT-5.2-Thinking (45.5).
The model also excels at coding tasks. On Arena-Hard v2, it achieved a score of 90.2, significantly outperforming competitors such as Claude-Opus-4.5 (76.7).
An affordable alternative for developers
Alibaba Cloud has positioned Qwen3-Max-Thinking as a premium yet affordable offering. Pricing is as follows:
- Input: CZK 30 per 1 million tokens (for standard contexts ≤ 32k)
- Output: CZK 150 per 1 million tokens
This pricing structure is aggressive, undercutting many older flagship models while offering state-of-the-art performance. By comparison, GPT-5.2 costs approximately CZK 44 for input and CZK 350 for output per million tokens, while Claude Opus 4.5 charges CZK 125 for input and CZK 625 for output.
Developers should note the granular pricing for the new agentic capabilities. Alibaba separates the cost of "thinking" (tokens) from the cost of "action" (tool use):
- Agent Search Strategy: Both standard and advanced strategies are priced at CZK 250 per 1,000 calls
- Web Extractor: Free (limited-time offer)
- Code Interpreter: Free (limited-time offer)
Developer ecosystem compatibility
Alibaba Cloud has ensured that Qwen3-Max-Thinking is ready for immediate integration. The API supports the standard OpenAI format, allowing teams to switch models simply by changing the base_url and model name. In a smart move to capture the coding market, the API also supports the Anthropic protocol, making Qwen3-Max-Thinking compatible with Claude Code, a popular agentic coding environment.
The model is now available in Qwen Chat and through Alibaba Cloud Model Studio with an OpenAI-compatible HTTP API. Developers can begin experimenting with the model after registering an Alibaba Cloud account and creating an API key.
Sources: venturebeat.com and marktechpost.com
