Google yakaburitsa mhuri itsva yemamodheru emitauro mikuru, Gemma 4, zvichibva paGemini 3 model. Gemma 4 inogoverwa pasi peApache License mumhando dzakasiyana dzine maparamita 2.3, 4.5, 25.2, uye 30.7 bhiriyoni (E2B, E4B, 31B, uye 26B A4B). Mhando dzeE2B neE4B dzakakodzera kushandiswa pamafoni, masisitimu eInternet of Things (IoT), uye mabhodhi akafanana neRaspberry Pi, nepo dzimwe mhando dzakakodzera kushandiswa pamabasa ekushandira uye masisitimu ane maGPU evatengi. Saizi yemamiriro ezvinhu inotariswa nemodheru iyi i128 token yemamodheru eE2B neE4B, uye 256 token yemamodheru e31B ne26B A4B.
Mamodheru aya ane mitauro yakawanda uye ane modal dzakawanda: mitauro makumi matatu nemashanu inotsigirwa kubva mubhokisi (mitauro inopfuura 140 yakashandiswa panguva yekudzidziswa), uye zvinyorwa nemifananidzo zvinogona kugadziriswa sekupinda (modheru dzeE2B neE4B dzinotsigirawo kugadziriswa kweaudio). Modheru ye26B A4B yakavakirwa paMixture-of-Experts (MoE), umo modheru iyi yakakamurwa kuita nhevedzano yemanetiweki enyanzvi (kuburitsa mhinduro kunogona kushandisa maparamita mabhiriyoni 3.8 chete, asi kumhanya kwacho kwakanyanya kupfuura kwemodheru dzakakura dzekare), nepo dzimwe mhando dzichishandisa monolithic architecture yekare.
Mamodheru aya anotsigira nzira dzekufunga uye dzekugadzirisa pfungwa, uye anotsigira Basa reSystem rekugadzirisa mirairo (mitemo, zvirambidzo) zvakasiyana nedata. Mamodheru aya anogona kushandiswa pakunyora kodhi, kuziva zvinhu mumifananidzo, kuongorora vhidhiyo yefuremu nefuremu, kuongorora magwaro nemaPDF, kuziva hunhu hwemavara (OCR) ezvinyorwa zvakadhindwa nezvakanyorwa nemaoko, kuziva kutaura, uye kushandura pakati pemitauro. Anogonawo kushandiswa seanoshanda ega achishandisa maturusi akasiyana-siyana nemaAPI.
Mumiedzo yakawanda, mamodheru eGemma 4 akabudirira zvakanyanya kupfuura mamodheru eGemma 3 ane maparamita mabhiriyoni makumi maviri nemanomwe. Gemma 4 inotsigira LiteRT-LM, vLLM, llama.cpp, MLX, Ollama, NVIDIA NIM neNeMo, LM Studio, Unsloth, SGLang, Cactus, Basetan, MaxText, Tunix, naKeras.
Source: opennet.ru
