Alibaba releases an ‘open’ challenger to OpenAI’s o1 reasoning model

Alibaba releases QwQ-32B-Preview, a reasoning AI rival to OpenAI's o1.

: Alibaba has unveiled QwQ-32B-Preview, a reasoning AI model designed to compete with OpenAI's o1. This model, developed by the Qwen team, boasts 32.5 billion parameters and surpasses o1 in certain benchmarks like AIME and MATH. While effective in logic puzzles and math, it can produce language errors and has limits on common sense reasoning. The model is available on Hugging Face under the Apache 2.0 license, but only some components are disclosed.

Alibaba has introduced QwQ-32B-Preview, an AI model aimed at challenging OpenAI's o1 in reasoning capabilities. Created by Alibaba's Qwen team, this model contains 32.5 billion parameters and can process prompts of up to 32,000 words, outstripping o1 in selective tests like AIME and MATH.

QwQ-32B-Preview is proficient in solving logic puzzles and responding to complex math queries, though it occasionally encounters issues with language switching and tasks involving common sense. Unlike most AI models, it includes self-fact-checking abilities, allowing it to sidestep common errors, albeit at a slower pace.

The model is accessible via Hugging Face, although full replication is hindered due to incomplete component release. Amidst advances in this area, Google and other organizations are investing heavily in reasoning models, questioning the effectiveness of merely increasing data and compute power under traditional scaling laws.

AI modelQwQ32BPreviewReasoning AI