{"id":125519,"date":"2025-05-23T15:05:12","date_gmt":"2025-05-23T13:05:12","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/mistral-ai-opublikoval-devstral-bolshuyu-yazykovuyu-model-dlya-raboty-s-kodom"},"modified":"2025-05-23T15:05:12","modified_gmt":"2025-05-23T13:05:12","slug":"mistral-ai-opublikoval-devstral-bolshuyu-yazykovuyu-model-dlya-raboty-s-kodom","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/news\/mistral-ai-opublikoval-devstral-bolshuyu-yazykovuyu-model-dlya-raboty-s-kodom","title":{"rendered":"Mistral AI publikoi Devstral, nj\u00eb model t\u00eb madh gjuh\u00ebsor p\u00ebr pun\u00ebn me kodin","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Kompania Mistral AI prezantoi nj\u00eb model t\u00eb madh gjuh\u00ebsor Devstral, t\u00eb optimizuar p\u00ebr t\u00eb zgjidhur problemet q\u00eb dalin gjat\u00eb zhvillimit t\u00eb softuerit. N\u00eb ndryshim nga modelet standarde t\u00eb AI, Mistral AI del jasht\u00eb shkrimit t\u00eb funksioneve t\u00eb ve\u00e7anta dhe kompletimeve t\u00eb kodit, duke ofruar mund\u00ebsi p\u00ebr t\u00eb analizuar dhe kontekstualizuar (p\u00ebrcaktuar q\u00ebllimin dhe logjik\u00ebn e funksionimit) bazat e m\u00ebdha t\u00eb kodit, p\u00ebrcaktuar lidhjet mes komponenteve dhe identifikuar gabimet e v\u00ebshtira p\u00ebr t'u zbuluar n\u00eb funksione t\u00eb nd\u00ebrlikuara. <\/p>\n
Modeli p\u00ebrfshin 23.6 miliard parametra, merr parasysh kontekstin n\u00eb 128 mij\u00eb tokene dhe \u00ebsht\u00eb publikuar n\u00ebn licenc\u00ebn Apache 2.0. Arkivi i shkarkuesh\u00ebm i Devstral z\u00eb 47 GB dhe \u00ebsht\u00eb i p\u00ebrshtatsh\u00ebm p\u00ebr p\u00ebrdorim n\u00eb sisteme lokale \u2014 p\u00ebr t\u00eb ekzekutuar modelin mjafton nj\u00eb PC me kart\u00ebn grafike NVIDIA GeForce RTX 4090 dhe 32 GB RAM. Modeli mund t\u00eb p\u00ebrdoret n\u00eb mjetet e hapura SWE-agent dhe OpenHands p\u00ebr automatizimin e korrigjimeve t\u00eb gabimeve, analiz\u00ebn e kodit dhe p\u00ebrmbushjen e ndryshimeve. <\/p>\n
Sistemi mund t\u00eb p\u00ebrdoret p\u00ebr t\u00eb zgjidhur probleme specifike (issue) n\u00eb GitHub dhe tejkalon ndjesh\u00ebm projektet e tjera n\u00eb grupin e testit SWE-Bench Verified, i cili kontrollon sakt\u00ebsin\u00eb e zgjidhjeve t\u00eb problemeve tipike n\u00eb kod (ofrohen 500 teste bazuar n\u00eb mesazhe reale gabimesh n\u00eb GitHub). N\u00eb k\u00ebt\u00eb test, modeli Devstral arriti 46.8%, nd\u00ebrsa modeli Claude 3.5 Haiku mori 40.6%, SWE-smith-LM 32B \u2014 40.2% dhe GPT-4.1-mini \u2014 23.6%. P\u00ebrve\u00e7 k\u00ebsaj, Devstral tejkaloi modele t\u00eb m\u00ebdha si Deepseek-V3-0324 671B (38.8%) dhe Qwen3 232B-A22B (34.4%), q\u00eb p\u00ebrfshijn\u00eb qindra miliarda parametra. \n Burimi: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"