Qwen3‑Max‑Thinking Boosts Performance with Test‑Time Scaling—Why It Still Isn’t Open‑Source

Overview

Qwen3‑Max‑Thinking is presented as the “full‑version” of Qwen3‑Max, shifting focus from post‑training improvements to inference‑time enhancements. The model now performs adaptive tool calls—searching the web when needed and invoking a Python interpreter for calculations—without explicit prompting.

Key Features

Adaptive Tool Calls : The model automatically decides whether to perform a web search or run Python code during reasoning.

Multi‑turn Self‑Reflection : It can look back at earlier reasoning steps, accumulate experience, and correct mistakes on the fly.

The official term for this approach is “Experience‑Cumulative Test‑Time Scaling Strategy (TTS)”, which means the model spends more time on a question to improve accuracy, becoming smarter with each attempt.

Benchmark Improvements

Mathematics : Score on IMO‑AnswerBench rose from 83.9 to 91.5 after enabling TTS.

Code Generation : LiveCodeBench v6 score reached 91.4, closing the previously noted coding gap.

Agent Capability : HLE (with tools) score jumped to 58.3, indicating stronger tool‑use proficiency.

Integration

The model is fully compatible with the OpenAI API, allowing existing LangChain, Dify, One‑API, and similar ecosystems to switch by changing the base URL and model name, resulting in near‑zero migration cost.

Model ID:

qwen3-max-2026-01-23

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/api/v2/apps/protocols/compatible-mode/v1",
)

response = client.responses.create(
    model="qwen3-max-2026-01-23",
    input="你好，请介绍一下 Qwen3-Max-Thinking 有什么新特性？",
)
print(response.output_text)

The client.responses.create endpoint is a proprietary DashScope interface that supports richer outputs such as search results and chain‑of‑thought reasoning, though the standard Chat Completions endpoint also works.

Thinking Process

The core of Qwen3‑Max‑Thinking is its “Thinking” capability, which internalizes adaptive tool use. During inference, the model decides whether to fetch up‑to‑date information or write Python code to verify a mathematical conjecture, embodying a System 2 (slow‑thinking) approach that the author likens to a student allowed to use calculators and dictionaries during an exam.

Conclusion

From a technical standpoint, Qwen3‑Max‑Thinking directly addresses the logical reasoning and complex task‑calling shortcomings of its predecessor. Its strengths include markedly stronger reasoning, seamless tool integration, and excellent OpenAI‑compatible compatibility.

Pros : Superior inference ability, smooth tool calls, easy migration.

Potential Concerns : Higher token consumption and longer latency due to the Thinking mode; early‑stage API stability may require limited testing.

For users disappointed with the original Qwen3‑Max, the Thinking version offers a compelling upgrade, delivering an AI that can reflect and use tools—closer to the envisioned “super assistant”.