Topics
Latest
AI
Amazon
Image Credits:Yuchiro Chino(opens in a new window)/ Getty Images
Apps
Biotech & Health
mood
Image Credits:Yuchiro Chino(opens in a new window)/ Getty Images
Cloud Computing
Department of Commerce
Crypto
Enterprise
EVs
Fintech
Fundraising
Gadgets
back
Government & Policy
Hardware
layoff
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
Security
Social
Space
startup
TikTok
Transportation
Venture
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
newssheet
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
Chinese house continue to relinquish AI models that rival the capabilities of system developed by OpenAI and other U.S.-based AI companies .
This workweek , MiniMax , an Alibaba- and Tencent - bet on inauguration that hasraisedaround $ 850 million in venture capital and is appraise at more than $ 2.5 billion , debutedthreenew models : MiniMax - Text-01 , MiniMax - VL-01 , and T2A-01 - HD . MiniMax - Text-01 is a text - only simulation , while MiniMax - VL-01 can see both paradigm and textbook . T2A-01 - HD , meanwhile , generates audio — specifically language .
MiniMax lay claim that MiniMax - Text-01 , which is 456 billion parameters in size , perform better than model such as Google ’s recently unveiledGemini 2.0 Flashon benchmarks like MMLU and SimpleQA , which measure the ability of a mannequin to resolve maths problems and fact - establish questions . Parameters roughly correspond to a model ’s problem - solve skills , and models with more parameters generally perform better than those with fewer parameter .
As for MiniMax - VL-01 , MiniMax says that it rivals Anthropic’sClaude 3.5 Sonneton evaluations that require multimodal intellect , like ChartQA , which task models with answering graph- and diagram - link up queries ( for example , “ What is the vizor time value of the orange line in this graphical record ? ” ) . grant , MiniMax - VL-01 does n’t quite best Gemini 2.0 Flash on many of these tests . OpenAI’sGPT-4oand an heart-to-heart model called InternVL2.5 beat it on several as well .
Of note , MiniMax - Text-01 has an extremely large context windowpane . A model ’s context , or context window , refers to input ( e.g. , text ) that a model see before generating output ( additional text ) . With a context window of 4 million token , MiniMax - Text-01 can analyze around 3 million lyric in one go — or just over five copy of “ War and Peace . ”
For setting ( no pun designate ) , MiniMax - Text-01 ’s setting windowpane is roughly 31 multiplication the size of it of GPT-4o ’s and Llama 3.1 ’s .
The last of MiniMax ’s mannequin released this week , T2A-01 - HD , is an audio generator optimise for speech . T2A-01 - HD can generate a synthetical voice with adjustable cadence , tone , and tenor in around 17 different languages , include English and Chinese , and clone a voice from just 10 seconds of an audio recording .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
MiniMax did n’t publish bench mark results compare T2A-01 - HD to other audio - bring forth poser . But to this newsman ’s auricle , T2A-01 - HD ’s outputs sound on par with audio frequency example fromMetaand startups likePlayAI .
With the exception of T2A-01 - HD , which is solely useable through MiniMax ’s API and Hailuo AI platform , MiniMax ’s new models can be download from GitHub and the AI dev chopine Hugging Face .
Just because the models are “ openly ” available does n’t mean they are n’t locked down in certain aspects , however . MiniMax - Text-01 and MiniMax - VL-01aren’t genuinely open sourcein the mother wit that MiniMax has n’t release the component part ( for instance , grooming data ) needed to re - create them from scratch . Moreover , they ’re under MiniMax ’s restrictive license , which prohibits developer from using the modelling to improve rival AI role model and requires that platform with more than 100 million monthly active users request a special permit from MiniMax .
MiniMax was found in 2021 by former employees of SenseTime , one of China ’s large AI firms . The troupe ’s project include apps like Talkie , an AI - powered role - playing platform along the lines ofCharacter AI , and text - to - television models that MiniMax has released in Hailuo .
Some of MiniMax ’s products have become the subject of venial contention .
Talkie , which was pulled from Apple ’s App Store in December for unspecified “ technical ” reasons , feature AI embodiment of public figures , let in Donald Trump , Taylor Swift , Elon Musk , and LeBron James , none of whom seem to have consented to being featured in the app .
In December , Broadcast magazinereportedthat MiniMax ’s video generator can reproduce the logos of British television TV channel , suggesting that MiniMax ’s models were trained on content from those channels . And MiniMax is reportedlybeing suedby iQiyi , a Taiwanese video streaming service that say MiniMax illicitly trained on iQiyi ’s copyright transcription .
MiniMax ’s newfangled model get in day after the outgoing Biden administrationproposedharsher export rules and restriction on AI engineering for Taiwanese ventures . company in China were already foreclose from buying advanced AI chips , but if the new rule go into burden as indite , company will be faced with stricter caps on both the semiconductor unit tech and models needed to bootstrap sophisticated AI system .
On Wednesday , the Biden administrationannouncedadditional standard focalize on keeping advanced microprocessor chip out of China . Chip foundries and promotional material companionship that want to export certain cow dung will be subjected to all-embracing license requirements unless they exercise greater scrutiny and due app to forbid their merchandise from reaching Taiwanese client .
