On the morning of April 29, Alibaba open-sourced the new-generation Qwen3 model of Tongyi Qwen (referred to as Qwen3), with a parameter count only one-third that of DeepSeek-R1, significantly reducing costs while comprehensively outperforming models like R1 and OpenAI-o1, ranking it as the strongest open-source model globally.

Qwen3 is the first "hybrid reasoning model" domestically, integrating "fast thinking" with "slow thinking" into the same model, greatly saving on computational resources.

According to official statements, the flagship version Qwen3-235B-A22B achieved the same level as DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro in benchmark tests for code, math, and general capabilities.

In the AIME25 Olympiad-level assessment, Qwen3-235B-A22B scored 81.5, setting a new record for open-source models; in the LiveCodeBench evaluation testing code abilities, Qwen3-235B-A22B broke through 70 points, surpassing Grok 3; in the ArenaHard evaluation assessing model human preference alignment, Qwen3-235B-A22B scored 95.6, surpassing OpenAI-o1 and DeepSeek-R1.

Under the same computational resources, the Qwen3 model surpasses larger-scale previous-generation models with smaller scale, truly achieving "small but powerful."

The total parameter count of Qwen3 is 235 billion, setting a new high for the intelligence level of open-source models. Alibaba stated that Qwen3's full-strength version can be deployed with just four H20s, with memory usage being only one-third that of similar-performance models.

The Qwen3 model versions include two MoE models of 30 billion and 235 billion parameters, as well as six dense models of 0.6 billion, 1.7 billion, 4 billion, 8 billion, 14 billion, and 32 billion parameters.

Meanwhile, Qwen3 provides better support for the upcoming AI agent and large model application boom. In the BFCL evaluation assessing model agent capabilities, Qwen3 set a new high score of 70.8, surpassing top models like Gemini2.5-Pro and OpenAI-o1, significantly lowering the threshold for agent tool invocation.

It is reported that the Qwen3 series models continue to use the permissive Apache2.0 license for open-source purposes, supporting 119+ languages for the first time. Developers, research institutions, and enterprises worldwide can download and commercialize the models for free from platforms such as ModelScope and HuggingFace, or call the Qwen3 API service via Alibaba Cloud Bailian. Individual users can immediately experience Qwen3 directly through the Tongyi app, and Quark will soon fully integrate Qwen3.

Currently, Alibaba's Tongyi has open-sourced over 200 models, with global downloads exceeding 300 million times. The number of derivative models from Qwen exceeds 100,000, surpassing Llama from the U.S., becoming the world's leading open-source model.

This article is an exclusive piece by Observer Network and unauthorized reproduction is prohibited.

Original source: https://www.toutiao.com/article/7498570054523437620/

Disclaimer: This article represents the author's personal views. Please express your opinions by clicking the "Upvote/Downvote" buttons below.