Alibaba’s newly unveiled open-source reasoning model, QwQ-32b (Quan-with-Questions), poses a potential challenge to established players like DeepSeek and OpenAI. With fewer parameters and lower resource requirements, QwQ-32b promises performance comparable to DeepSeek’s R1 and OpenAI’s models. This open-source model is readily available for experimentation, though with some limitations.
DeepSeek AI running on an iPhone.
DeepSeek’s R1 large language model (LLM) gained attention in February for its cost-effective performance rivaling ChatGPT. Alibaba’s QwQ-32b aims to push the boundaries further. DeepSeek-R1 utilizes 671 billion parameters, with 37 billion activated, while QwQ-32b operates with only 32 billion. This translates to a significant difference in computational demands: DeepSeek R1 requires 1600GB of VRAM, whereas QwQ-32b functions with just 24GB, compatible with high-end GPUs like Nvidia’s H100 or even the RTX 4090 and 5090.
Accessing QwQ-32b through Qwen Chat
Available under an Apache 2.0 license, QwQ-32b is accessible to companies, researchers, and the public through Alibaba’s Qwen Chat platform. Like DeepSeek, Qwen Chat has limitations but offers some immediate advantages.
Qwen Chat: In-Depth Answers and Reasoning
One notable feature is Qwen Chat’s tendency to provide detailed responses, even to simple queries. While offering extensive context, this can sometimes feel excessive. A positive aspect is the transparency of its reasoning process, reminiscent of ChatGPT’s Deep Thinking, albeit with less depth.
Limitations and Potential
When questioned about political topics, Qwen Chat flags them as inappropriate. While potential jailbreak methods might exist, similar to DeepSeek, they haven’t been readily apparent. The validity of Alibaba’s performance claims remains to be thoroughly tested.
A New Competitor in the LLM Arena
QwQ-32b’s emergence presents a potential shift in the LLM landscape. With its reduced computational requirements and open-source nature, it positions itself as a competitor to established models like ChatGPT and DeepSeek. Whether it truly matches their performance remains to be seen, but QwQ-32b marks a significant development in accessible and potentially cost-effective LLM technology.
