Alibaba unveiled the signboard model ‘QWen 3’. Although there isn’t any outstanding innovation function, it reflects some great benefits of the recent flagship models, and claims that it exceeds the most recent models of Google and Open AI in some performance.
Alibaba unveiled ‘Q1 3’ on the twenty eighth (local time). It consists of varied sizes starting from 600 million to 235 billion parameters.
Two of them have applied a MOE method to 2 models. One is ‘Q13-235B-A22B’ with a complete of 235 billion parameters and 22 billion energetic parameters, and the opposite is a small model ‘Q13-30B-A3B’ with a complete of 30 billion parameters and three billion energetic parameters.
Nevertheless, it only unveiled the benchmark for this model and didn’t launch an actual model.
As a substitute, it was released as an open source. These six models should not MOE, but Dense models that utilize your complete parameters.
In any case, Alibaba was the largest feature of Q13, which was tolerated through the MOE architecture, dividing the work, sharing the work into knowledgeable model, and increasing the computation efficiency. This is similar architecture as ‘Deep Chic-R1’.
It is usually a ‘hybrid’ model that may select reasoning-viscosity. In complex problems, time and pondering processes are used to infer and respond quickly to easy requests. This method has been the usual for brand new models, with the introduction of ‘Rock-3’ for the reason that launch of ‘Claude 3.7 Sonnet’ by Antropic.
It supports a complete of 119 languages and learned about 36 trillion tokens. Learning data consists of textbooks, Q & A pairs, code, and AI production contents.
As well as, the coding and agent functions were optimized, and the model context protocol (MCP) support was also enhanced. This can also be the identical trend.
Consequently of the benchmark, the general performance has been improved over the previous version, Q12. In some tests, it showed higher results from Open AI’s O3-Mini or Google’s ‘Geminai 2.5 Pro’.
Specifically, the massive MOE model, Q13-235B-A22B, surpassed the O3-mini in ‘Codeforces’, Mathematics Test ‘AIME’, and BFCL. As well as, the small MOE model, Q1 3-30B-A3B, has a greater performance than the ‘QWQ-32B’, which has 10 times more energetic parameters.
The biggest model currently released is Q13-32B, which is competitive in comparison with several open source and industrial models, including ‘Deep Chic-R1’. It scored the next rating in LiveCodebench than Open AI ‘O1’.
The Q -One 3 -line up shows similar performance to the Q1 2.5 lineup, which had more parameters.
In other words, Q1 3-1.7b/4b/8b/14B/32B model shows performance to correspond to the cue one 2.5-3b/7b/14b/32b/72b model. Specifically, it’s ahead of Q1 2.5 in STEM, coding, and reasoning.
He also emphasized that he has excellent functions as an agent, akin to tool-calling, instructions, and copying specific data formats.
Q13 is Huband GitHubIt would be available or soon provided by open license through open license, and the community of Alibaba Cloud ModelScopeIt will probably even be utilized in.
By Park Chan, reporter cpark@aitimes.com
