Alibaba’s Qwen Team Releases QwQ-32B Open-Source Reasoning Model, Said to Perform Similar to DeepSeek-R1

Alibaba’s Qwen Crew, a division tasked with creating artificial intelligence (AI) fashions, launched the QwQ-32B AI mannequin on Wednesday. It’s a reasoning mannequin primarily based on prolonged check time compute with seen chain-of-thought (CoT). The builders declare that regardless of being smaller in measurement in comparison with the DeepSeek-R1, the mannequin can match its efficiency primarily based on benchmark scores. Like different AI fashions launched by the Qwen Crew, the QwQ-32B can be an open-source AI mannequin, nonetheless, it isn’t absolutely open-sourced.

QwQ-32B Reasoning AI Mannequin Launched

In a blog post, Alibaba’s Qwen Crew detailed the QwQ-32B reasoning mannequin. QwQ (brief for Qwen with Questions) collection AI fashions had been first launched by the corporate in November 2024. These reasoning fashions had been designed to supply an open-source various for the likes of OpenAI’s o1 collection. The QwQ-32B is a 32 billion parameter mannequin developed by scaling reinforcement studying (RL) methods.

Explaining the coaching course of, the builders mentioned that the RL scaling strategy was added to a cold-start checkpoint. Initially, RL was used just for coding and mathematics-related duties, and the responses had been verified to make sure accuracy. Later the method was used for normal capabilities together with rule-based verifiers. The Qwen Crew discovered that this methodology elevated normal capabilities of the mannequin with out lowering its math and coding efficiency.

QwQ-32B AI Mannequin benchmarks
Photograph Credit score: Alibaba

The builders declare that these coaching constructions enabled the QwQ-32B to carry out at comparable ranges to the DeepSeek-R1 regardless of the latter being a 671-billion-parameter mannequin (with 37 billion activated). Primarily based on inner testing, the group claimed that QwQ-32B outperforms DeepSeek-R1 within the LiveBench (coding), IFEval (chat or instruction fine-tuned language), and the Berkeley Perform Calling Leaderboard V3 or BFCL (means to name capabilities) benchmarks.

Builders and AI lovers can discover the open weights of the mannequin on Hugging Face itemizing and Modelscope. The mannequin is accessible underneath the Apache 2.0 licence which permits tutorial and research-related utilization however forbids industrial use circumstances. Moreover, because the full coaching particulars and datasets should not obtainable, the mannequin can be not replicable or may be deconstructed. DeepSeek-R1 was additionally obtainable underneath the identical licence.

In case one lacks the appropriate {hardware} to run the AI model domestically, they will additionally entry its capabilities through Qwen Chat. The mannequin picker menu on the top-left of the web page will let customers choose the QwQ-32B-preview mannequin.