Qwen3-Max-Thinking Explained: Adaptive Tool Use, Experience-Based Test-Time Scaling and API Access

In late January 2026, Alibaba’s Qwen team released Qwen3-Max-Thinking, its flagship reasoning model. It builds on Qwen3-Max, which Alibaba has described as having more than one trillion parameters and being pre-trained on 36 trillion tokens. The Thinking version adds two things: tools the model can call on its own, and a new test-time scaling method for a “heavy” mode. This article explains both, the published results, how to call the API, and what to weigh before using it.

What Alibaba announced

According to the official Qwen blog post, Qwen3-Max-Thinking was trained with substantial reinforcement-learning compute and improves on factual knowledge, complex reasoning, instruction following, alignment with human preferences and agent tasks. Across 19 established benchmarks, the team says its performance is comparable to GPT-5.2-Thinking, Claude Opus 4.5 and Gemini 3 Pro.

Note that, unlike many Qwen models, Qwen3-Max-Thinking is not an open-weight release. It is available through Qwen Chat and Alibaba Cloud’s Model Studio API.

Adaptive tool use

Earlier tool-using chat models often required users to switch tools on before a task. Qwen3-Max-Thinking chooses by itself when to use its built-in Search, Memory and Code Interpreter during a conversation. The team says this came from fine-tuning for tool use followed by further training with rule-based and model-based feedback, and that Search and Memory help reduce hallucinations, provide current information and personalise answers, while the Code Interpreter handles computation.

In practice this matters most in Qwen Chat, where the tools are built in. In the API, developers still decide which capabilities to expose and how to handle results.

Test-time scaling with “take-experience”

Test-time scaling means spending more computation at inference to get better answers. A common approach is to sample many reasoning paths in parallel and aggregate them, for example by majority vote. The Qwen team argues this often produces redundant reasoning.

Instead, the heavy mode uses an experience-cumulative, multi-round strategy. It limits the number of parallel trajectories and spends the saved compute on rounds of self-reflection. A “take-experience” mechanism distils key insights from earlier rounds, so later rounds do not re-derive known conclusions and focus on what is still uncertain. Because the model carries forward distilled insights rather than raw transcripts, it uses its context window more efficiently.

The blog reports gains over standard parallel sampling at roughly the same token consumption:

Benchmark	Standard parallel sampling	Experience-cumulative
GPQA	90.3	92.8
HLE	34.1	36.5
LiveCodeBench v6	88.0	91.4
IMO-AnswerBench	89.5	91.5
HLE (with tools)	55.8	58.3

The widely cited 58.3 on Humanity’s Last Exam is the with-tools score in heavy mode. The team compares it with 45.5 for GPT-5.2-Thinking and 45.8 for Gemini 3 Pro. It also says it blocked access to Hugging Face and other HLE-related websites during evaluation to prevent answer leakage. Without tools, the heavy-mode HLE score is 36.5, so comparisons should match settings.

Using the API

The API is available through Alibaba Cloud Model Studio. The blog shows two routes.

OpenAI-compatible endpoint:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)
completion = client.chat.completions.create(
    model="qwen3-max-2026-01-23",
    messages=[{"role": "user", "content": "Give me a short introduction to large language models."}],
)
print(completion.choices[0].message.content)

Anthropic-compatible endpoint, which lets tools such as Claude Code use Qwen3-Max-Thinking by setting ANTHROPIC_BASE_URL to Alibaba’s endpoint, ANTHROPIC_MODEL to qwen3-max-2026-01-23 and the auth token to a DashScope API key.

Keep API keys in environment variables or a secrets manager rather than in code.

Trade-offs to consider

• Latency and cost. Thinking modes, and especially multi-round heavy modes, generate many more tokens than a direct answer. They suit complex analysis, coding and research, not quick chat replies. Check current Model Studio pricing before large-scale use.
• Data location. Requests go to Alibaba Cloud endpoints; organisations with data-residency requirements should check which region they use.
• Vendor benchmarks. The results are from the Qwen team. Independent evaluations and your own tests on real tasks are the better guide.
• Opacity of reasoning. Multi-round self-reflection can make it harder to see why the model reached a conclusion. Asking for sources and verifying key facts remains important.

Where it fits

Qwen3-Max-Thinking is aimed at hard reasoning and agentic tasks where accuracy is worth extra latency. Its API compatibility with OpenAI and Anthropic formats lowers switching costs, which makes it practical to include in a multi-model setup: route simple queries to cheaper models and send the hardest problems to a thinking model.

Sources

• Alibaba Cloud Community: Pushing Qwen3-Max-Thinking Beyond its Limits
• Qwen blog: Qwen3-Max-Thinking