Topics
up-to-the-minute
AI
Amazon
Image Credits:Yuichiro Chino / Getty Images
Apps
Biotech & Health
Climate
Image Credits:Yuichiro Chino / Getty Images
Cloud Computing
mercantilism
Crypto
Alibaba’s internal benchmark results for Qwen3.Image Credits:Alibaba
endeavour
EVs
Fintech
Fundraising
Gadgets
gage
Government & Policy
Hardware
layoff
Media & Entertainment
Meta
Microsoft
privateness
Robotics
Security
Social
blank space
inauguration
TikTok
Transportation
Venture
More from TechCrunch
event
Startup Battlefield
StrictlyVC
newssheet
Podcasts
telecasting
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
Chinese tech company Alibaba on MondayreleasedQwen3 , a family of AI models that the company claims can match and , in some cases , outperform the good models available from Google and OpenAI .
Most of the models are — or soon will be — uncommitted for download under an “ opened ” permit on AI dev platformHugging FaceandGitHub . They range in sizing from 0.6 billion parameters to 235 billion parameter . ( parametric quantity approximately correspond to a model ’s trouble - solve acquisition , and models with more argument mostly perform better than those with few parameters . )
The rise of China - originated model series like Qwen have increased the pressure on American science laboratory such as OpenAI to cede more equal to AI engineering . They ’ve also lead policymakers to follow through restrictions get at limiting the power of Chinese AI companies to obtain thechipsnecessaryto gear models .
introduce Qwen3 !
We eject and open - free weight Qwen3 , our latest large lyric models , include 2 MoE framework and 6 heavy good example , roam from 0.6B to 235B. Our flagship framework , Qwen3 - 235B - A22B , reach competitive consequence in benchmark evaluations of put one across , math , general…pic.twitter.com/JWZkJeHWhC
— Qwen ( @Alibaba_Qwen)April 28 , 2025
According to Alibaba , the Qwen3 modelling are “ hybrid ” models — they can take time to “ reason ” through complex trouble , or answer simple requests quickly . Reasoning turn on the model to in effect fact - check themselves , similar to model like OpenAI’so3 , but at the price of high latency .
“ We have seamlessly incorporate thinking and non - intelligent way , offering users the flexibility to manipulate the thinking budget , ” the Qwen team wrote in ablog post . “ This design enables users to configure task - specific budgets with greater simpleness . ”
Some of the framework also adopt a smorgasbord of experts ( MoE ) computer architecture , which can be more computationally effective for answer queries . MoE breaks down tasks into subtasks and designate them to smaller , specialized “ expert ” framework .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
The Qwen3 models support 119 language , Alibaba said , and were school on a dataset of over 36 trillion tokens . ( Tokens are the raw spot of data that a model outgrowth ; 1 million tokens is equivalent to about 750,000 Scripture . ) The company sound out Qwen3 was train on a combination of textbooks , “ question - response dyad , ” code snip , AI - generated information , and more .
These improvements , along with others , greatly supercharge Qwen3 ’s capabilities compared to its predecessor , Qwen2 , Alibaba said . None of the Qwen3 good example seem to be drumhead and shoulder above the top - of - the - personal line of credit recent example like OpenAI ’s o3 and o4 - mini , but they ’re strong performers all the same .
On Codeforces , a platform for programming contest , the largest Qwen3 exemplar — Qwen-3 - 235B - A22B — just beat OpenAI’so3 - miniand Google’sGemini 2.5 Pro . Qwen-3 - 235B - A22B also scoop o3 - mini on the latest version of AIME , a ambitious math bench mark , and BFCL , a examination for assessing a model ’s ability to “ reason ” about problems .
But Qwen-3 - 235B - A22B is n’t publicly available — at least not yet .
The largest public Qwen3 mannequin , Qwen3 - 32B , is still private-enterprise with a routine of proprietary and undetermined AI manakin , including Chinese AI laboratory DeepSeek’sR1 . Qwen3 - 32B surpasses OpenAI’so1model on several tests , include the coding bench mark LiveCodeBench .
Alibaba said Qwen3 “ excels ” in tool - forebode capabilities as well as following instructions and copy specific data formats . In addition to the models for download , Qwen3 is available from cloud providers , including Fireworks AI and Hyperbolic .
Tuhin Srivastava , Centennial State - founder and CEO of AI cloud host Baseten , allege Qwen3 is another gunpoint in the trend line of unfastened model keeping yard with closed - source scheme such as OpenAI ’s .
“ The U.S. is doubling down on trammel sales of chip to China and purchase from China , but models like Qwen 3 that are body politic - of - the - artistic creation and open [ … ] will undoubtedly be used domestically , ” he told TechCrunch . “ It reflects the reality that businesses are both progress their own tools [ as well as ] buying off the shelf via shut - model companies like Anthropic and OpenAI . ”
