On April 29, 2025, Chinese tech giant Alibaba released and open-sourced its new-generation Qwen3 large language model (LLM). So, how does this new model perform?
According to public information, the predecessor of Qwen3 — the Qwen1.5-110B model — performed outstandingly in multiple benchmark tests and is comparable to Meta’s Llama3-70B model.
In this update, Qwen3 introduces a hybrid reasoning mechanism that combines the generative capabilities of traditional language models with dynamic reasoning abilities. This enhances the model’s performance on complex tasks, especially those requiring logical reasoning and multi-step derivations. Multiple evaluations indicate that Qwen3 outperforms top-tier models like DeepSeek-R1 and OpenAI-o1 across several benchmarks, including coding, mathematics, and general capabilities, making it arguably the strongest open-source LLM available today.
So, what highlights of Qwen3 are worth getting excited about?
Multilingual Support & Cross-Platform Compatibility: In addition to Chinese and English, Qwen3 now supports over 29 languages, including German, Japanese, and French. This demonstrates a significant improvement in cross-lingual capabilities, and Qwen3 also boasts excellent cross-platform compatibility, making it easy to deploy across different operating systems and hardware environments.
Hybrid Reasoning Capabilities: Qwen3 introduces “thinking mode” and “non-thinking mode.” It merges traditional LLM generation with dynamic reasoning, enhancing performance in complex scenarios, particularly those involving logical inference and step-by-step problem solving.
Power of Open Source: Alibaba has open-sourced the entire Qwen3 series (including flagship models) under the Apache 2.0 license. This move is highly significant for the democratization and development of AI technology.
High Cost-Performance Ratio: Qwen3 delivers top-tier performance while drastically reducing deployment costs. For example, the flagship Qwen3-235B-A22B model only requires 4 H20 GPUs for deployment, whereas similar-performance models like DeepSeek-R1 may require 8 or more. Additionally, Huawei’s computing division announced that its Ascend MindSpeed and MindIE platforms have long supported the Qwen series and now offer out-of-the-box support for Qwen3, achieving 0-day compatibility upon open-sourcing.
After Alibaba released Qwen3, Tesla founder Elon Musk immediately posted on X:
Next week, Grok 3.5 early beta release to SuperGrok subscribers only. It is the first AI that can, for example, accurately answer technical questions about rocket engines or electrochemistry.
So what surprises might Grok 3.5 bring?
Coincidentally, also today, Meta hosted its first-ever “LlamaCon” conference. This event marks Meta’s inaugural AI developer gathering, aimed at further promoting its open-source Llama models, attracting more developers to contribute to the open-source AI ecosystem, and collectively advancing generative AI technology.
Also today, Duolingo — with hundreds of millions of users — announced its transformation into an “AI-first” company. The company is embracing AI across the board, including in recruitment, performance evaluation, and workforce planning, to enhance operational efficiency.
Artificial intelligence continues to exert increasing influence across the globe. Whether tech giants or rising stars, no one wants to miss this wave. So, who will ultimately emerge as the winner?
Creating this content wasn’t easy — for more quality articles about AI, please visit: https://iaiseek.com/news