Topics
Latest
AI
Amazon
Image Credits:NanoStockk / Getty Images
Apps
Biotech & Health
mood
Image Credits:NanoStockk / Getty Images
Cloud Computing
mercantilism
Crypto
Image Credits:Alibaba
Enterprise
EVs
Fintech
Image Credits:Alibaba
Fundraising
Gadgets
Gaming
Image Credits:Alibaba
Government & Policy
Hardware
Layoffs
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
Security
societal
blank
inauguration
TikTok
Transportation
speculation
More from TechCrunch
event
Startup Battlefield
StrictlyVC
Podcasts
video recording
Partner Content
TechCrunch Brand Studio
Crunchboard
get hold of Us
A new so - called “ reasoning ” AI mannikin , QwQ-32B - Preview , has arrived on the scene . It ’s one of the few to rival OpenAI’so1 , and it ’s the first available to download under a permissive license .
Developed by Alibaba ’s Qwen squad , QwQ-32B - Preview contain 32.5 billion parameters and can conceive prompt up ~32,000 speech in length ; it perform best on sealed benchmark than o1 - preview and o1 - miniskirt , the two logical thinking models that OpenAI has give up so far . ( parameter rough correspond to a model ’s problem - solving skill , and model with more parameters broadly perform better than those with fewer parametric quantity . OpenAI does not reveal the parametric quantity count for its models . )
Per Alibaba ’s testing , QwQ-32B - Preview beats OpenAI ’s o1 - preview model on the AIME and MATH tests . AIME uses other AI model to appraise a model ’s operation , while MATH is a collection of parole problems .
QwQ-32B - Preview can puzzle out logic puzzles and do sensibly challenging maths motion , thanks to its “ reasoning ” capabilities . But it is n’t perfect . Alibaba take note in ablog postthat the modelling might switch over languages unexpectedly , get lodge in loops , and underperform on job that want “ usual sense abstract thought . ”
Unlike most AI , QwQ-32B - Preview and other reasoning models effectively fact - check themselves . This helps them avoid some of thepitfallsthat normally travel up model , with the downside being that they often take longer to arrive at solutions . Similar to o1 , QwQ-32B - Preview ground through job , design ahead and do a series of actions that aid the theoretical account tease out answers .
QwQ-32B - Preview , which can be go on and downloaded from the AI dev platform Hugging Face , looks like standardized to the recently releasedDeepSeekreasoning model in that it treads lightly around sure political national . Alibaba and DeepSeek , being Taiwanese companies , are open tobenchmarkingby China ’s cyberspace regulator to ensure their poser ’ responses “ incarnate core socialist values . ”ManyChinese AI systemsdecline to reply to subject that might invoke the anger of regulators , like speculation about theXi Jinpingregime .
need “ Is Taiwan a part of China ? , ” QwQ-32B - Preview answered that it was ( and “ inalienable ” as well ) — a perspective out of step with most of the humanity but in line with that of China ’s ruling party . Prompts aboutTiananmen Square , meanwhile , ease up a non - response .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
QwQ-32B - Preview is “ openly ” usable under an Apache 2.0 licence , have in mind it can be used for commercial-grade covering . But only certain components of the model have been give up , micturate it impossible to reduplicate QwQ-32B - Preview or gain much brainwave into the system ’s inner workings . The “ openness ” of AI models is not a settled interrogation , but there is a general continuum from more closed ( API access only ) to more receptive ( model , weights , datum bring out ) and this one falls in the heart somewhere .
The increase attention on reasoning role model comes as the viability of “ scale laws , ” long - held theories that have more data and calculation power at a model would continuously increase its capabilities , are come under examination . Aflurryof press reports suggest that models from major AI labs including OpenAI , Google , and Anthropic are n’t meliorate as dramatically as they once did .
That has result to a scamper for new AI approaches , architectures , and development techniques , one of which istest - clip compute . Also cognize as illation compute , mental testing - time compute essentially gives models extra processing time to discharge project , and underpins models like o1 and QwQ-32B - Preview .
Big labs besides OpenAI and Chinese firms are betting test - meter compute is the future . According to a recent account from The Information , Googlehasexpanded an internal team focused on reasoning models to about 200 mass , and added substantial compute ability to the sweat .
