Google yatulutsa banja latsopano la mitundu yayikulu ya zilankhulo, Gemma 4, kutengera mtundu wa Gemini 3. Gemma 4 imagawidwa pansi pa Apache License m'mitundu yosiyanasiyana yokhala ndi magawo 2.3, 4.5, 25.2, ndi 30.7 biliyoni (E2B, E4B, 31B, ndi 26B A4B). Mitundu ya E2B ndi E4B ndi yoyenera kugwiritsidwa ntchito pazida zam'manja, machitidwe a Internet of Things (IoT), ndi ma Raspberry Pi-like board, pomwe mitundu ina ndi yoyenera kugwiritsidwa ntchito pa malo ogwirira ntchito ndi machitidwe okhala ndi ma GPU a ogula. Kukula kwa mawonekedwe komwe kumaganiziridwa ndi chitsanzochi ndi ma tokeni 128 a mitundu ya E2B ndi E4B, ndi ma tokeni 256 a mitundu ya 31B ndi 26B A4B.
Ma model amenewa ndi a zilankhulo zambiri komanso amitundu yosiyanasiyana: Zilankhulo 35 zimathandizidwa kuchokera m'bokosi (zilankhulo zoposa 140 zinagwiritsidwa ntchito panthawi yophunzira), ndipo zolemba ndi zithunzi zimatha kukonzedwa ngati zolowera (ma model a E2B ndi E4B amathandiziranso kukonza mawu). Ma model a 26B A4B amachokera ku kapangidwe ka Mixture-of-Experts (MoE), momwe model imagawidwa m'magulu angapo a akatswiri (kupanga mayankho kungagwiritse ntchito magawo 3.8 biliyoni okha, koma liwiro lake ndi lalikulu kwambiri kuposa la ma model akuluakulu akale), pomwe mitundu ina imagwiritsa ntchito kapangidwe ka monolithic kakale.
Ma model amathandizira njira zoganizira komanso zosinthira zomwe zingasinthidwe, ndipo amathandizira Udindo Wadongosolo pokonza malangizo (malamulo, zoletsa) padera pa deta. Ma model angagwiritsidwe ntchito polemba ma code, kuzindikira zinthu pazithunzi, kusanthula kanema wa chimango ndi chimango, kusanthula zikalata ndi ma PDF, kuzindikira zilembo (OCR) za mawu osindikizidwa ndi olembedwa pamanja, kuzindikira mawu, ndi kumasulira pakati pa zilankhulo. Angagwiritsidwenso ntchito ngati othandizira odziyimira pawokha omwe amalumikizana ndi zida zosiyanasiyana ndi ma API.
Mu mayeso ambiri, mitundu ya Gemma 4 yachita bwino kwambiri kuposa mtundu wa Gemma 3 wa magawo 27 biliyoni. Gemma 4 imathandizira LiteRT-LM, vLLM, llama.cpp, MLX, Ollama, NVIDIA NIM ndi NeMo, LM Studio, Unsloth, SGLang, Cactus, Basetan, MaxText, Tunix, ndi Keras.
Source: opennet.ru
