Alibaba drops an open challenger to OpenAI's o1

Alibaba's Qwen team just dropped QwQ-32B-Preview, an open-source reasoning model aimed directly at OpenAI's o1.

Instead of rushing to output the first token, QwQ spends extra test-time compute to plan ahead, talk through difficult problems step by step, and fact-check its own logic before delivering an answer.

What it does and how it performs

QwQ-32B-Preview packs 32.5 billion parameters and handles prompt contexts up to 32,000 words (32K tokens).

According to Alibaba's benchmarks, the model beats OpenAI's o1-preview on challenging math evaluations like AIME and MATH, and outpaces o1-mini across several key reasoning and programming tests.

It excels at multi-step logic puzzles and math questions where conventional models tend to hallucinate.

That said, it is still an early preview with clear rough edges. Alibaba openly notes that the model occasionally switches languages mid-thought, gets stuck in repetitive reasoning loops, and underperforms on simple common-sense questions.

Like other models developed in China, it also declines to answer certain politically sensitive queries.

Specs, licensing, and where to get it

QwQ-32B-Preview is hosted on Hugging Face and released under an Apache 2.0 license, making it available for commercial use.

At 32.5B parameters, it is practical enough for developers to run locally or fine-tune with reasonable GPU hardware, especially when quantized.

Keep in mind that while the model weights are accessible, Alibaba hasn't released the full training data or recipe, placing it squarely in the semi-open category rather than full reproducibility.

Why it matters

If you are building AI agents, coding assistants, or complex math workflows, QwQ-32B-Preview is definitely worth spinning up. Until now, o1-style deep introspection was locked behind closed proprietary APIs; having a 32B model that you can self-host and experiment with changes that dynamic completely.

While OpenAI's models remain more polished on language coherence and common sense, QwQ shows that open-source labs are mastering inference-time compute at record speed, shrinking the frontier reasoning moat faster than expected.