After OpenAI released models with breakthrough reasoning capabilities, the competition for ai reasoning capabilities has begun, with alibaba and hengfang sequentially launching significant new products that not only match the performance of the o1 model but are also open source!

On Thursday, alibaba's Tongyi Qianwen launched the QwQ-32B-Preview open-source model, which includes 32.5 billion parameters and can handle prompts of up to 32,000 tokens. It outperforms OpenAI's reasoning models o1-preview and o1-mini in AIME and MATH benchmark tests.

QwQ is one of the few models that can compete with o1, excelling in mathematics and programming, especially in complex problems requiring deep reasoning, and it can be used for commercial applications.

Last week, the algo giant hengfang's DeepSeek-R1-Lite model preview surpassed o1-preview on challenging mathematics and coding tasks, significantly outpacing GPT-4o and others. In the AIME test benchmark, its score steadily increased with computation time.

It is noteworthy that the officials also stated that the model is still in development and, through continuous iterations, the official version of the DeepSeek-R1 model will be fully open-sourced.

The emergence of alibaba and hengfang's models signals the rise of reasoning ai in the industry, which may provide opportunities for small ai developers to catch up, breaking the current dominance of a few technology giants.

Fireworks, a startup that began researching reasoning models in the second quarter of this year, has its co-founder and CEO Lin Qiao stating:

The entire open-source community... will release inference models at a super fast pace.

In addition, technology giants are also increasing their efforts in the development of inference models. Google has expanded the scale of its inference model team from a few dozen people before the o1-preview release to around 200 people, and Google has also provided more computing resources for this team.

Latecomers have more cost advantages; the thinking chain is key to large models.

Latecomers have more cost advantages in building large models.

Latecomers seem to benefit from the research papers on inference published in recent years by Stanford University, Google, meta platforms, and OpenAI's own researchers when developing alternatives to OpenAI. The development cost of inference models is lower than that of traditional LLMs, such as GPT-4o, which require hundreds of millions of dollars on computational resources and training data, and must legally obtain that data.

The new models can help OpenAI and its competitors develop coding assistants capable of completing difficult projects. For example, enterprise software companies like microsoft and Salesforce can use them to improve agents that take action on behalf of clients, such as scheduling appointments.

It is worth mentioning that researchers can incorporate reasoning capabilities into existing LLMs by having other models generate the thinking processes for solving problems, which can then be used to train LLMs.

Some researchers have also made inference-focused datasets available for free to other developers. For example, alibaba stated that it used data from one of Open o1's research teams to construct its inference model.

Ion Stoica, the co-founder of the ai startup Anyscale and Databricks, stated:

In developing inference models, competitors of OpenAI do not have a clear disadvantage.

Editor/Lambor