By Keshavan Thiagaraja
Kwanaki biyu kacal bayan Google ta sanar da cewa samfurinta mafi karfi zai kasance ga kungiyar masu kare yanar gizo da aka zaba kawai, Mistral AI ta dauki hanyar da take akasin haka. A ranar 29 ga Oktoba, dakin gwaje-gwajen da ke Paris ya fitar da weights na Mistral Large 4, babban samfurin ta mai triliyan daya na sigogi, domin kowa ya iya saukar da shi, bincika shi kuma ya kula da shi a kan na’urorin nasa.
Lambobi na bayyana dalilin da yasa wannan abu ne mai muhimmanci. Bayan karin makonni uku na reinforcement learning tun daga lokacin da aka fara duba shi a bainar jama’a, Mistral Large 4 yanzu yana jagorantar benchmark ta FrontierCode 2.0 ta Cognition da kashi 19.7, da Terminal-Bench 4.0 da kashi 41, maki goma sha uku gaban preview nasa yayin da yake amfani da output tokens kashi talatin da fi karancinsu. A kan Artificial Analysis Coding Agent Index, ya yi daidai da jagoran samfuran open-weight, GLM-5.3, da kashi 48, yayin da yake amfani da kashi 47 na tokens a kowace aiki a matsakaici. A cikin kimantawar Surge Code da mutane suka yi ba tare da sanin sunaye ba, masu kimantawa sun fi son shi bisa GLM-5.3, Kimi K3 da DeepSeek V4.1 Flash, kuma sun sanya shi a bayan Claude Opus 5.5 kacal.
Labarin tsaron ma yana da tsauri kamar haka. Mistral Large 4 ya samu kashi 51 a kan Artificial Analysis Cyber Index, wanda ke gwada ko samfuri na iya gano raunuka a cikin codebase, sake samar da su da kuma gyara su ba tare ya lalata kome ba. Yana aiwatar da wadannan ayyukan kariya ba tare da kin yarda da ba a bukata ba, yayin da yake kin yarda da kashi 96.6 na bukatun yanar gizo marasa kyau, mafi girman yawan da Mistral ta taba auna a cikin samfuran da aka bude. Booz Allen Hamilton, Cisco, Cloudflare, Fortinet, Rubrik da Zscaler duk sun goyi bayan wannan fitowar.
Wannan ita ce irin wannan kwarewar da Google ta yanke shawarar rarrabawa ta hanyar takaita. Wani dakin gwaje-gwaje ya kammala da cewa samfurin da ya yi kwarewa wajen gano da gyara raunuka dole ne ya kasance a bayan bincike da kwangila. Wancan ya kammala da cewa ana bauta wa masu kariya mafi kyau lokacin da kowa zai iya rike makami daya. Dukkan hujjoji suna da nauyi, kuma masana’antu yanzu za su kalli ana gwada su tare a fili a gaban kowa.
Lambobin kamfanoni suna cika zancen. Mistral Large 4 ya kai kashi 73.3 a kan benchmark na ayyukan kuɗi Finch, kashi 84.5 a kan BrowseComp na binciken yanar gizo, kuma yana jure kashi 97.1 na hare-hare a kan benchmark na B3 Agent Security, sakamako mafi kyau a cikin samfuran open-weight. Hanyoyin tunani guda uku, ƙarami, matsakaici da babba, suna ba ƙungiyoyin damar daidaita tsada da daidaito a kowace aiki.
Hujjar ikon mallaka ita ce motsin da yake da zurfi. An horar da samfurin daga farko a kan kayayyakin aikin Turai da asali bayyananne, kuma Mistral tana faɗin masu sauraronta a bayyane: gwamnatoci, bankuna da masana’antu masu ka’idoji waɗanda ba za su iya aika bayanan su zuwa gajimare na wani ba. La Banque Postale, wata banki ta gwamnati, ta ce ikon fasaha ba zabin ba ne wajen kiyaye amanar abokan ciniki. Haka kuma ana samun wani nau’in da aka duba na tsaron yanar gizo tare da abokan hulɗa kamar Orange Cyberdefense a ƙarƙashin kariyar kwangila.
Ga shugabanni, tambayar aikin ta canza. Tambayar ba shine ko samfurin da aka bude yana da kyau ga aikin gaske ba, domin wannan a bayyane yake yana da shi. Tambayar ita ce abin da ƙungiyarku za ta yi da samfurin gaba da ta ke iya sarrafa shi gaba ɗaya: ta yi masa bincike, ta keɓance shi da bayanan ku, da gudanar da shi inda dokokin ku ke aiki, ko kuwa ta ci gaba da hayar iyawa ta sharuɗɗan wasu.
Bude ƙofar da taka-tsantsan duk suna a kan allon yanzu. Ƙungiyoyin da suka koyi yau yadda ake kimanta, sarrafa da kuma shigar da samfuran open-weight sune za su kafa nasu sharuɗɗa gobe, ko wane falsafa zai yi nasara.
Yi tunani a lissafi. Yi aiki da wayo. Jagoranta da dabara. Gina da alhakin amfani. Fadada a duniya. Kirkira ainihin darajar kasuwanci ta hanyar basira wucin gadi.
#ArtificialIntelligence #AIStrategy #MachineLearning #Research #ITImpress


