![\"\"](\"https:\/\/ekonomigundemi.com.tr\/wp-content\/uploads\/2026\/08\/4e44b867f7dedad994866416550caf2b_1.jpg\")
\n<\/p>\nDijital varl\u0131k sekt\u00f6r\u00fcn\u00fcn en b\u00fcy\u00fck \u015firketi olan Tether\u2019in yapay zek\u00e2 ara\u015ft\u0131rma giri\u015fimi QVAC, mobil cihazlar ve u\u00e7 cihazlarda (edge) \u00e7al\u0131\u015fmak \u00fczere geli\u015ftirilen VisionPsy-Nano adl\u0131 g\u00f6r\u00fcnt\u00fc ve dil modelini (Vision-Language Model – VLM) a\u00e7\u0131k kaynak olarak yay\u0131mlad\u0131\u011f\u0131n\u0131 duyurdu.<\/p>\n
Yakla\u015f\u0131k 460 milyon parametreye sahip VisionPsy-Nano, daha \u00f6nce bulut altyap\u0131s\u0131 gerektiren \u00e7ok modlu (multimodal) yapay zek\u00e2 yeteneklerini do\u011frudan mobil cihazlara ta\u015f\u0131yor. Model, 0,5 milyar parametrenin alt\u0131ndaki cihaz \u00fczerinde \u00e7al\u0131\u015fan g\u00f6r\u00fcnt\u00fc ve dil modelleri aras\u0131nda 62,3 normalize puanla kalite ve performans s\u0131ralamas\u0131nda ilk s\u0131raya yerle\u015fti.<\/p>\n
Kompakt yap\u0131s\u0131yla sekt\u00f6r testlerinde \u00f6ne \u00e7\u0131kt\u0131<\/p>\n
VisionPsy-Nano, yakla\u015f\u0131k 0,5 milyar parametre s\u0131n\u0131f\u0131ndaki modeller aras\u0131nda yap\u0131lan kar\u015f\u0131la\u015ft\u0131rmalarda Liquid AI ve Hugging Face taraf\u0131ndan geli\u015ftirilen rakip modelleri geride b\u0131rakt\u0131. Yap\u0131lan testlerde VisionPsy-Nano-460M; Liquid AI\u2019\u0131n LFM2.5-VL-450M, Hugging Face\u2019in SmolVLM2-500M ve temel modeli olan nanoVLM-460M-8k ile kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131. Model, 17 de\u011ferlendirme kriterinin 16\u2019s\u0131nda en y\u00fcksek performans\u0131 g\u00f6stererek 62,3 normalize puana ula\u015ft\u0131.<\/p>\n
Farkl\u0131 kullan\u0131m ihtiya\u00e7lar\u0131na y\u00f6nelik iki s\u00fcr\u00fcm sunuyor<\/p>\n
Tether, VisionPsy-Nano\u2019yu farkl\u0131 kullan\u0131m senaryolar\u0131na uygun iki ayr\u0131 s\u00fcr\u00fcmle geli\u015ftiricilerin kullan\u0131m\u0131na sunuyor. VisionPsy-Nano-460M, kalite ve do\u011fruluk odakl\u0131 yap\u0131s\u0131yla standart g\u00f6r\u00fcnt\u00fc ve dil testlerinde s\u0131n\u0131f\u0131n\u0131n en y\u00fcksek performans\u0131n\u0131 sunuyor. VisionPsy-Nano-460M-Flash ise mobil cihazlarda \u00e7ok daha d\u00fc\u015f\u00fck gecikmeyle \u00e7al\u0131\u015facak \u015fekilde optimize edildi. Model, tam s\u00fcr\u00fcm\u00fcn performans\u0131n\u0131n yakla\u015f\u0131k y\u00fczde 99\u2019unu korurken, ak\u0131ll\u0131 telefonlarda \u00e7ok daha h\u0131zl\u0131 yan\u0131t verebiliyor.<\/p>\n
Mobil cihazlarda geli\u015fmi\u015f yapay zek\u00e2 deneyimi sunuyor<\/p>\n
Kompakt yap\u0131s\u0131na ra\u011fmen VisionPsy-Nano bir\u00e7ok farkl\u0131 alanda y\u00fcksek performans sergiliyor. Model; karma\u015f\u0131k belgelerden metin \u00e7\u0131karma (OCR), finansal raporlar ve infografikleri analiz etme, sahne alg\u0131lama ve mek\u00e2nsal analiz, g\u00f6rsel muhakeme ile komutlar\u0131 do\u011fru yorumlama gibi g\u00f6revlerde kendi kategorisindeki en ba\u015far\u0131l\u0131 sonu\u00e7lara ula\u015ft\u0131. Ayr\u0131ca MM-IFEval ve POPE gibi de\u011ferlendirme testlerinde, kendisinden 1,6 ila 2,3 kat daha b\u00fcy\u00fck modelleri geride b\u0131rakarak dikkat \u00e7ekici bir ba\u015far\u0131 elde etti.<\/p>\n
Ak\u0131ll\u0131 telefonlarda \u00e7ok daha h\u0131zl\u0131 \u00e7al\u0131\u015f\u0131yor<\/p>\n
VisionPsy-Nano-460M-Flash, daha az g\u00f6rsel veri i\u015fleyerek ilk yan\u0131t\u0131 \u00fcretme s\u00fcresini \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131yor. Yap\u0131lan testlerde model; Pixel 9, Galaxy S23 ve Galaxy S25 Ultra cihazlar\u0131nda SmolVLM2-500M ve temel modele k\u0131yasla yakla\u015f\u0131k 19 ila 23 kat, iPhone 15’te ise 36 kata kadar daha h\u0131zl\u0131 ilk yan\u0131t \u00fcretebildi. Ayn\u0131 zamanda gecikme s\u00fcresi bak\u0131m\u0131ndan LFM2.5-VL-450M ve Qwen3.5-0.8B modellerinden de daha iyi performans g\u00f6sterdi.<\/p>\n
Her iki model de Apache 2.0 lisans\u0131 alt\u0131nda a\u00e7\u0131k a\u011f\u0131rl\u0131k (open-weight) olarak yay\u0131mland\u0131. Ara\u015ft\u0131rmac\u0131lar ve geli\u015ftiriciler; Hugging Face Transformers, llama.cpp tabanl\u0131 mobil \u00e7al\u0131\u015ft\u0131rma altyap\u0131s\u0131 ve \u00fcretim ortamlar\u0131na y\u00f6nelik vLLM deste\u011fi \u00fczerinden modellere eri\u015febilecek. Ayr\u0131ca t\u00fcm test yap\u0131land\u0131rmalar\u0131 ve de\u011ferlendirme s\u00fcre\u00e7leri de sonu\u00e7lar\u0131n yeniden \u00fcretilebilmesini sa\u011flamak amac\u0131yla a\u00e7\u0131k olarak payla\u015f\u0131l\u0131yor.<\/p>\n
\u201cYapay zek\u00e2y\u0131 veri merkezlerinden cihazlara ta\u015f\u0131yoruz\u201d<\/p>\n
Tether CEO\u2019su Paolo Ardoino konuya ili\u015fkin yapt\u0131\u011f\u0131 a\u00e7\u0131klamada, \u201cTeknoloji art\u0131k sekt\u00f6r\u00fcn veri merkezlerine ba\u011f\u0131ml\u0131 yapay zek\u00e2 yakla\u015f\u0131m\u0131na alternatif sunabilecek olgunlu\u011fa ula\u015ft\u0131. Sadece 460 milyon parametreyle s\u0131n\u0131f\u0131n\u0131n en iyi kalite ve performans\u0131n\u0131 elde etmemiz, yerel \u00e7al\u0131\u015fan ve y\u00fcksek verimlilik sunan yapay zek\u00e2n\u0131n g\u00fc\u00e7l\u00fc bir alternatif oldu\u011funu g\u00f6steriyor. Bu ara\u00e7lar\u0131 do\u011frudan geli\u015ftiricilerin kullan\u0131m\u0131na a\u00e7arak merkezi platformlara olan ba\u011f\u0131ml\u0131l\u0131\u011f\u0131 azalt\u0131yor, g\u00fc\u00e7l\u00fc ve gizlili\u011fi koruyan yapay zek\u00e2y\u0131 insanlar\u0131n halihaz\u0131rda kulland\u0131\u011f\u0131 cihazlara ta\u015f\u0131yoruz.\u201d dedi.<\/p>\n
\n<\/p>\n
\u00a0<\/p>\n
<\/strong> Tether Data\u2019n\u0131n yapay zek\u00e2 ara\u015ft\u0131rma giri\u015fimi QVAC, mobil cihazlarda \u00e7al\u0131\u015fmak \u00fczere geli\u015ftirilen VisionPsy-Nano isimli g\u00f6r\u00fcnt\u00fc ve dil modelini a\u00e7\u0131k kaynak olarak kullan\u0131ma sundu. Yakla\u015f\u0131k 460 milyon parametreye sahip model, kendi kategorisinde kalite ve performans a\u00e7\u0131s\u0131ndan sekt\u00f6r testlerinde ilk s\u0131rada yer ald\u0131.<\/p>\n","protected":false},"author":1,"featured_media":126911,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"googlesitekit_rrm_CAow07ekDA:productID":"","footnotes":""},"categories":[1,9],"tags":[],"class_list":["post-126910","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ekonomi","category-manset"],"_links":{"self":[{"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/posts\/126910","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/comments?post=126910"}],"version-history":[{"count":0,"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/posts\/126910\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/media\/126911"}],"wp:attachment":[{"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/media?parent=126910"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/categories?post=126910"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ekonomigundemi.com.tr\/index.php\/wp-json\/wp\/v2\/tags?post=126910"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
\nEKONOM\u0130 GAZETES\u0130 <\/p>\n","protected":false},"excerpt":{"rendered":"