Alibaba has released a new AI reasoning model through which the company is openly challenging OpenAI’s o1.
OpenAI recently launched the OpenAI o1, which can fact-check itself. The new product, codenamed Strawberry, is a family of AI models that help the client-end AI models execute self-reasoning to provide more accurate outputs. Alibaba has launched its in-house reasoning AI model, QwQ-32B-Preview, pitting it directly with OpenAI’s o1-Preview and o1-Mini.
The new AI model is available to download under a permissible license and is trained on 32.5 billion parameters. It can easily comprehend and work on prompts of up to 32000 words in length and outperforms the benchmarks set by o1-preview and o1-mini.
Alibaba’s AI model also beats OpenAI’s reasoning models in AIME and MATH tests. AIME test uses other AI models to evaluate the tested model’s performance, and MATH focuses on a collection of word problems. Alibaba’s newly launched product also solves logic puzzles and provides answers to reasonably challenging mathematical problems.
However, the flagship AI model comes with its cons during the initial launch, which the company claims to rectify in the fastest timelines possible. They include switching languages abruptly, getting stuck in loops, and underperforming on tasks related to common sense reasoning. What sets the model apart is its ability to fact-check its answers and correct them during the generation without causing delays in the output.
QwQ-32B-Preview can be downloaded from the AI Dev platform Hugging Face. Alibaba’s new AI model behaves similarly to other Chinese-produced AI models in adhering to the region’s core socialist values. When asked if Taiwan is a part of China, the model confirms that it is inalienable, an added exaggeration bent towards the Chinese ruling party’s perspective.
QwQ-32B-Preview comes with an Apache 2.0 license, meaning it can be used commercially.
Though the general source code is kept open, certain components like API access have been kept closed, making the product’s replication nearly impossible. This will make it challenging for its closest competitors to identify the source of capabilities and give Alibaba a longer competitive edge.
