Chii chaitika
PrismML yakaburitsa Bonsai 2 27B, chimiro chemutauro chakakumisikidzwa chakagadzirirwa kushanda pamakomputa emunhu uye mafoni epamusoro-soro asina makore. Iyo modhi inoderedza Alibaba's Qwen3.8 27B kusvika 5.9 gigabytes uchishandisa ternary uremu compression, ichichengeta 98% yeyekutanga performance. Huremu huripo pasi peApache 2.0 License uye inotsigira NVIDIA GPUs kuburikidza neCUDA uye Apple zvishandiso kuburikidza neMLX.
PrismML, yekutanga yakavambwa neCaltech vaongorori uye inotungamirwa naBabak Hassibi, yakaburitsa Bonsai 2 27B nemusi weChina. Iyo modhi ishanduro yakamisikidzwa yeAlibaba's Qwen3.8 27B, yakaderedzwa kusvika 5.9 gigabytes kuburikidza neternary uremu compression, iyo inoshandura 16-bit parameters kune matatu kukosha: yakanaka, yakaipa imwe, kana zero. Iyi tekinoroji inobvisa zvisingaite ma paramita, ichibvumira iyo modhi kuchengetedza 98% yeyekutanga bhenji rekuita panguva yakakodzera pane yemuno Hardware.
Kuburitswa kwacho kunoratidza kuvandudzwa kwakakosha pamusoro peiyo yekutanga Bonsai modhi kubva munaKurume, iyo yakachengeta 95% yebenchmark performance. Sekureva kwaPrismML, iyo yekutanga modhi yakadhindwa kanopfuura mamirioni gumi nerimwe, iine mamwe madiki mamodheru anounganidza 2.6 miriyoni yekudhawunirodha, zvichiratidza kugadzwa kwakasimba kwemugadziri wemudziyo-based AI. Iyo modhi nyowani inomhanya paNVIDIA GPUs kuburikidza neCUDA uye paApple zvishandiso (Mac, iPhone, iPad) kuburikidza neMLX, ichishandisa tsika yakaderera-bit kernels kuve nechokwadi chekuita kwemuno.
Iwo maremu emhando anowanikwa nekukasira pasi peApache 2.0 License, inobvumira kubatanidzwa kwakafara uye kugadziridzwa. PrismML yakasimudza 22.25 miriyoni yembeu yembeu yakatsigirwa naKhosla Ventures, Cerberus Capital, uye Caltech, naIon Stoica, Databricks co-muvambi, anoshanda sechipangamazano. Maitiro ekambani anotarisa pakukwidziridza kudzokororwa mumamodhi akakura kuti awane yakakwira performance parity ine saizi shoma yefaira, zano reHassibi rinoratidza kuti richawedzera kushanda sezvo saizi dzemhando dzichiwedzera.
Kwakabva mashoko: techjuice.pk ↗
Nei zvichikosha
Kuburitswa uku kunoratidza kuti kufunga-inokwanisa AI inogona kushanda munharaunda pane yevatengi Hardware, kugadzirisa zvakavanzika uye kubvisa cloud latency uye kubhadhara mari. Nekuwana yakakwira performance parity ine diki diki faira saizi, PrismML inofambisira mberi shanduko kubva pagore-inotsamira AI kuenda kugoverwa, mudziyo-wakavakirwa kufunga, ichiita epamusoro AI kugona kuwanikwa pasina hukuru hwedata data.
Iko kugona kumhanyisa kufunga-inokwanisa AI munharaunda pamidziyo yevatengi inogadzirisa zvakakomba zvekuvanzika nekuona kuti data remushandisi harimbosiya mudziyo. Izvi zvinobvisa kudiwa kwekufambisa kwegore, kuderedza latency uye kubvisa mari inoenderera mberi yekunyorera ine hukama ne cloud-based AI services. Kune vanogadzira uye vashandisi vekupedzisira, izvi zvinomiririra shanduko yakakosha kuenda kune yakagoverwa, mudziyo-yakavakirwa kufunga iyo inodhura-inoshanda uye yakavanzika.
Iyo yakakwirira yekuchengeta mwero we98% performance mu5.9GB faira inoratidza kuti nzira dzekudzvanya dzinogona kuchengetedza hungwaru hwemhando huru. Izvi zvine zvazvinoita kumaindasitiri anoda pa-mudziyo AI, senge hutano, mari, uye vabatsiri vega, uko hutongi hwedata uye mhinduro chaiyo-nguva yakakosha. Iyo yakavhurika-sosi hunhu hwekuburitswa inowedzera kukurumidza kugamuchirwa nekubvumira nharaunda kuvakira pamusoro uye kukwidziridza modhi yezvimwe zvipingamupinyi zvehardware.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Vagadziri vanofanirwa kutarisa kuwanikwa kwetsika yakaderera-bit kernels yeakasiyana magadzirirwo emahara uye chinangwa chakataurwa chekambani chekumanikidza mazana-e-mabhiriyoni-parameter modhi mumwedzi iri kuuya. Pamusoro pezvo, iyo inogona kubatanidzwa neApple zvishandiso, inotaridzwa nenguva yekuburitswa, inobvumidza kutariswa kweramangwana-padivi AI kugona.
PrismML yakataura chinangwa chayo chinotevera kumanikidza modhi nemazana emabhiriyoni emaparamita mukati memwedzi mishoma inotevera. Kubudirira munzvimbo iyi kwaizowedzera kusimbisa kuwanda kweiyo ternary compression yemhando yemhando yemhando, zvichiita kuti hunyanzvi hweAI hwepamusoro huwanikwe pane Hardware yemuno.
Iyo nguva yekuburitswa yakonzera kufungidzira nezve inogona kubatanidzwa neApple zvishandiso. Nepo CEO Babak Hassibi vakaramba kutaura nezve mishumo yehurukuro neApple, kutsigirwa kweApple's MLX framework inoratidza kurongeka kwehurongwa neApple's ecosystem. Zviziviso zvenguva yemberi maererano nekudyidzana kwepamutemo kana mudziyo-chaiyo optimizations inogona kukanganisa zvakanyanya kutorwa kweAI yemuno pane iOS uye macOS mapuratifomu.