Topics
Latest
AI
Amazon
Image Credits:Jakub Porzycki/NurPhoto / Getty Images
Apps
Biotech & Health
Climate
Cloud Computing
Commerce
Crypto
Enterprise
EVs
Fintech
Fundraising
Gadgets
bet on
Government & Policy
computer hardware
layoff
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
Security
societal
blank
inauguration
TikTok
Transportation
speculation
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
video
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
Microsoft on Wednesdaylaunched several raw “ assailable ” AI models , the most capable of which is competitive with OpenAI ’s o3 - mini on at least one bench mark .
As it says on the tin can , all of the new permissively licensed modelling — Phi 4 mini logical thinking , Phi 4 reasoning , and Phi 4 reasoning plus — are “ reasoning ” models , intend they can spend more prison term fact - checking root to complex problem . They expand Microsoft ’s Phi “ belittled model ” family , which the company launched a twelvemonth ago to bid a base for AI developer building apps at the edge .
Phi 4 mini reasoning was trained on close to 1 million semisynthetic mathematics job generated by Chinese AI startup DeepSeek ’s R1 reasoning example . Around 3.8 billion parameters in size , Phi 4 mini abstract thought is design for educational applications , Microsoft say , like “ engraft tutoring ” on lightweight devices .
parametric quantity approximately check to a model ’s problem - solving skills , and models with more parametric quantity broadly execute better than those with fewer parameter .
Phi 4 reasoning , a 14 - billion - argument model , was trained using “ high - lineament ” web data as well as “ curated demonstrations ” from OpenAI ’s aforementioned o3 - miniskirt . It ’s best for mathematics , science and rally applications , according to Microsoft .
As for Phi 4 abstract thought plus , it ’s Microsoft ’s previously - released Phi 4 model adapted into a abstract thought example to reach good truth for particular task . Microsoft claim Phi 4 abstract thought plus approaches the execution storey of DeepSeek R1 , which has importantly more parameters ( 671 billion ) . The company ’s inner benchmarking also has Phi 4 abstract thought plus rival o3 - miniskirt on OmniMath , a math skills test .
Phi 4 miniskirt reasoning , Phi 4 reasoning , Phi 4 reasoning plus , and their detailed technical report , are available on theAI dev platform Hugging Face .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
“ Using distillate , reinforcement learnedness , and gamey - calibre data , these [ new ] modeling balance sizing and performance , ” compose Microsoft in ablog post . “ They are small enough for humble - latency environments yet assert strong abstract thought capabilities that rival much bigger models . This portmanteau allow even imagination - determine devices to perform complex logical thinking task efficiently . ”
