Alibaba launches QwQ, its open step-by-step reasoning model

Alibaba's Qwen team has unveiled QwQ-32B-Preview, a roughly 32-billion-parameter language model designed to solve complex math and coding problems by reasoning step by step before producing an answer. The company has released it as an open model, with weights available for download, letting developers and researchers run and study it without relying on a paid API.

What a reasoning model is

QwQ generates reasoning sequences before a conclusion, and its model card exposed that text to the user. A long explanation is not proof of correctness, and visible text should not automatically be called an “internal scratchpad”: the model still predicts tokens and can produce a plausible chain ending in a wrong answer. Evaluation must score the result and check every material step.

OpenAI had introduced o1-preview in September and said it would not expose the raw chain of thought to users. A useful comparison does not assume that both systems “think” in the same way; it fixes the task, compute budget, number of attempts, version and grading rule.

Alibaba's bet: open up the model

QwQ-32B-Preview, a name the company hasn't officially explained, builds on Alibaba's Qwen2.5 family and focuses specifically on this extended reasoning capability. The company has published it on Hugging Face under an open license, meaning anyone with enough hardware can download it, run it locally, and modify it—something no model in OpenAI's o1 family allows.

The name itself carries the "Preview" label: Alibaba is presenting the model as a preliminary version, not a finished product, and warns of limitations typical of a system still under development, such as possible reasoning loops or unexpected language switches mid-response. At 32 billion parameters, it's a mid-to-large-sized model, far from the scale of the biggest closed models but big enough to hold its own on math and coding benchmarks.

Why it matters

Until now, o1-style step-by-step reasoning had been almost exclusively OpenAI's turf. Alibaba releasing an open model aimed at the same problem changes the picture: it narrows the gap between what was only available through a closed, paid API and what any lab, university, or company can download and adapt at zero licensing cost. It also confirms that the "think before you answer" approach isn't a proprietary technique belonging to a single lab, but a research direction that several teams—including Chinese ones—are pursuing in parallel.

For the non-technical user, the practical difference is that this type of model tends to make fewer mistakes on problems requiring several chained logical steps—an equation with multiple unknowns, a function with several conditions—though at the cost of slower responses. For the industry, QwQ-32B-Preview is one more sign that the race for reasoning models is now being fought on two fronts: the closed systems built by major U.S. labs, and the open alternatives that Chinese companies like Alibaba are willing to put in anyone's hands.

Open weights do not mean zero cost

The QwQ card listed an Apache 2.0 licence, 32.5 billion parameters and a 32,768-token context. That allows the weights to be studied and adapted under the licence, but it does not remove memory, electricity, servers, maintenance or evaluation. “Free” can at most describe the absence of one licensing fee.

Model and product should also be separated. Weights do not necessarily include an interface, monitoring, filters, support or service guarantees. A company comparing a local model with an API must count hardware, concurrency, latency, staff and the cost of maintaining a secure version.

Testing reasoning without rewarding verbosity

Build a hidden set with verifiable answers, fix the number of samples and grade the conclusion before reading the explanation. Then classify errors: invented premise, calculation, code that does not run, omitted step or loop. The team itself warned about language mixing, recursive loops and weaknesses outside mathematics and coding.

Length needs a metric of its own. One system may be correct while using ten times more tokens; another may fail quickly. Time, cost, accuracy and verifiability form a vector, not one ranking. In code, running tests beats a persuasive explanation; in mathematics, substituting the result can expose errors hidden by a fluent chain.

A local reproduction recipe

A model name is not enough to repeat a result. Save the weight revision or hash, library version, chat template, system message, temperature, output limit and hardware. Preserve the exact question and grading rule as well. Changing any one of these can alter accuracy, length and speed.

Quantisation adds another variable. Lower-precision weights allow a model to run with less memory, but behaviour may change. A comparison should name the format and repeat a subset with the reference version. If only one variant fits the available machine, that constraint is part of the result rather than a technical detail to hide.

Visible reasoning, external verification

A long chain can be useful as an interface: it shows where to check, supports requests for another route and exposes some assumptions. It does not guarantee a faithful account of the process producing the answer. Evaluation therefore asks whether every intermediate claim is necessary, correct and compatible with the result, not whether it “sounds like thought.”

Tools change the examination. A calculator, code interpreter or search can correct model limitations, but the complete system is then being tested: which tool it chose, what data it sent, how it interpreted output and what it did after an error. Comparing a tool-equipped model with one lacking tools attributes an environmental advantage to the model.

Safe use begins by classifying the task. When an answer can be checked by a test, calculation or source, run that check. When health, money or rights are involved, model explanation does not replace a professional or authoritative document. When no verification route exists, uncertainty should remain visible rather than being filled with additional steps.

Benchmarks with conditions, not trophies

A score is comparable only when dataset, template, tools, sampling and grading rule match. Vendor results are hypotheses to reproduce, not guarantees for another task. A preview may also change while the product’s general name hides material differences.

Before accepting a benchmark, audit the published artifact as well: weight version, chat template, temperature, output limit and evaluation code. Changing one of those elements may alter the score while the model name stays unchanged. Saving the configuration and failed answers makes it possible to separate a system improvement from an advantage introduced by the protocol.

The transferable skill is to evaluate reasoning through traceability and result: exact version, budget, executable check and classified error. Visible steps can help locate a failure, but they do not turn the chain into evidence. External verification gets the final word.