MiniCPM-V 4.7 Uyasebenza

I-MiniCPM-V 4.7 iyimodeli enkulu ye-multimodal esebenza kumadivayisi, ekhishwe njenge-open source ngokubambisana yi-Mianbi Intelligence ne-OpenBMB, futhi inezinguqulo ezimbili: i-35B-A3B (MoE) ne-1B. Inguqulo ephambili iyimodeli yokuqala emhlabeni ye-open-source MoE multimodal esebenza kahle ngaphandle kokulahlekelwa ikhwalithi kumaselula; ngobuchwepheshe bokuhlela ochwepheshe obubheke phambili, idinga cishe u-3GB wememori futhi ingafinyelela ku-13+ tokens/s ku-iPhone, iholele amamodeli asezingeni elifanayo kuzo zonke izilinganiso ezingu-34; inguqulo ye-1B ihloselwe amadivayisi asezingeni lokuqala.

Ukuqonda okuvamile kwezithombe: ukubona izinto ezithombeni, ukuqhathanisa izithombe eziningi, nemibuzo nezimpendulo ezibonakalayo; amaphuzu angu-91.72 ku-MMBench ahola amamodeli asezingeni elifanayo.
Ukuhlaziya amadokhumenti ne-OCR: ukubona umbhalo, amathebula namagrafu ezithombeni noma kuma-PDF; amaphuzu angu-91.2 ku-OCRBench.
Ukucabanga ngezibalo ne-STEM: ukuphendula imibuzo yezibalo neyesayensi; MMMU 80.22, MathVista 88.
Ukuqonda amavidiyo amade: ukuqonda okuqukethwe kwamavidiyo anokulandelana kwesikhathi okude nokuphendula imibuzo ehlobene; amaphuzu angu-68.9 ku-LongVideoBench.
Ukusebenza kahle kumadivayisi: imodeli ye-35B idla cishe u-3GB wememori kuphela ku-iPhone futhi ikhiqiza ngesikhathi sangempela ngesivinini esingu-13+ tokens/s.
Ukulungisa nokwenza i-fine-tuning ngezindleko eziphansi: i-fine-tuning evulekile yokuzidlalela, efunda imisebenzi emisha ngaphandle kokukhohlwa amakhono amadala.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI open source.
Ukuhlela ochwepheshe okubheke phambili (LES): izisindo ze-MoE ezingu-65.6 GiB azikwazi ukufakwa zonke kumemori; ochwepheshe abaningi bagcinwa ku-SSD bese befundwa lapho kudingeka, kodwa ukuqondiswa kwesendlalelo ngasinye kuncike emiphumeleni yesendlalelo sangaphambilini, okwenza isendlalelo ngasinye silinde ukufundwa kwediski. I-LES isebenzisa ukufana okukhulu phakathi kokufakwayo kwezendlalelo eziseduze ukuze ibikezele kusenesikhathi ochwepheshe abazodingwa yisendlalelo esilandelayo futhi ibalayishe ngaphambili ngemuva; ngesikhathi esifanayo inikeza kuqala ukubalwa kochwepheshe asebelayishiwe nabakhethwe ukuqondiswa kwangempela. Lokhu kuvumela ukubala nokufunda kwenzeke ngesikhathi esisodwa, futhi izivivinyo zikhombisa ukusheshisa izikhathi ezingu-5 ngaphandle kokulahlekelwa ekuphumeni.
Ukuthuthukisa ukugcinwa nokufundwa kwezisindo: kusetshenziswa i-quantization yokunemba okuxubile ngesendlalelo ngasinye ukuze kuncishiswe usayizi wokugcina izisindo nenani ledatha edluliswayo; futhi ukuhlelwa kwezisindo zochwepheshe kumafayela kuyahlelwa kabusha, ukuze amapharamitha ayesakazeke ezindaweni ezintathu agcinwe ngokulandelana ngenombolo. Ngakho i-SSD ingafunda uchwepheshe ophelele ngokufunda ngokulandelana kanye, kuncishiswe kakhulu izindleko zokulayisha.
Ukucindezelwa kwama-visual token: kususelwa ku-LLaVA-UHD v4, ama-token acindezelwa kusenesikhathi ngaphakathi kwe-visual encoder, kuncishiswe ukubalwa okulandelayo kusukela emthonjeni; kusekelwa izinga lokucindezela elingu-4x noma 16x, futhi kumodi engu-16x izinkomba ezibalulekile cishe azilahlekelwa ikhwalithi, kugcinwe kokubili imininingwane yesithombe nokusebenzisana kwesikhathi sangempela.
I-DSpark speculative decoding: imojuli elula yokusalungisa ikhiqiza ama-token amaningi angaba khona ngesikhathi esisodwa, imodeli eyinhloko iwaqinisekise ngokuhambisana bese ikhipha amaningi ngesikhathi esisodwa. Ama-token amaningi ahlukanisa izindleko ezimisiwe zokufunda nokuhlela ochwepheshe be-MoE, futhi izivivinyo zibonisa ukuthi isivinini sokukhiqiza sikhuphuka ngokwesilinganiso izikhathi ezingu-1.17 ngaphandle komthelela kwikhwalithi.

Thola imodeli: landa izisindo zefayela lemodeli nekhodi ku-GitHub noma ku-Hugging Face.
Yizame ngokushesha endaweni: vakashela i-Web Demo esemthethweni ku-Hugging Face Spaces (MiniCPM-V-4.7-Demo) ukuze uyisebenzise ku-inthanethi ngaphandle kokufaka.
Faka uhlelo lokusebenza: vakashela indawo yokugcina i-MiniCPM-V-Apps ku-GitHub ukuze ulande amaklayenti e-iOS/Android, bese uwafaka ngokuqondile kuselula noma kuthebhulethi.
Khetha uhlaka lokuqagela lokuyisebenzisa: khetha eyodwa ku-llama.cpp, Ollama, MLX noma vLLM/SGLang ngokuya ngeplathifomu yehardware, ulayishe imodeli bese uqedela ukusethwa kwesevisi yokuqagela yasendaweni.
Vula ukusheshisa kwe-LES (ongakukhetha): lapho usebenzisa i-35B-A3B kumadivayisi, vula ukuhlela ochwepheshe okubheke phambili ngokwemibhalo ye-open source ukuze uthole ukusebenza okushelelayo ngememori engu-3GB.
Khetha imodi yokucindezela okubukwayo: shintsha phakathi kokucindezela okungu-4x noma okungu-16x ngokuya ngezidingo.

Indawo yokugcina ye-GitHub：https://github.com/OpenBMB/MiniCPM-V
Indawo yamamodeli ye-HuggingFace：https://huggingface.co/collections/openbmb/minicpm-v-47

Ukuqala emhlabeni kwe-MoE kumadivayisi: imodeli yokuqala emhlabeni ye-open-source MoE multimodal esebenza kumadivayisi ngaphandle kokulahlekelwa ikhwalithi futhi ngokushelela.
Ukusebenza kahle kwememori okwedlulele: imodeli ye-35B enezisindo ezingu-65.6 GiB idinga cishe u-3GB wememori kuphela, okuyingxenye engu-1/4 yememori yedivayisi engu-16GB.
Isivinini sokuqagela esiholayo: izivivinyo ku-iPhone 17 Pro Max zibonisa u-13+ tokens/s, kanti i-iPad Pro M4 ifinyelela ku-17+ tokens/s; ubuchwepheshe be-LES busheshisa ukusebenza izikhathi ezingu-5.
Ukusebenza kudlula amamodeli asezingeni elifanayo kuzo zonke izindawo: izilinganiso ezigunyaziwe ezingu-34 zihola i-Qwen3.6-35B-A3B, futhi idlula ngisho ne-Qwen 27B ne-Gemma-4 31B ezinkulu ngosayizi.
Ibhalansi engcono kakhulu phakathi kokusebenza nokusebenza kahle: ithola amaphuzu aphelele aphezulu ngesikhathi esifushane sokuqagela, ivuselele umngcele wokusebenza nokusebenza kahle kwe-multimodal kumadivayisi.
Izinhlobo ezimbili ezitholakala kabanzi: i-35B ephambili iphakamisa umkhawulo wamakhono, kanti i-1B enamandla ivumela ngisho namadivayisi asezingeni lokuqala ukuthi asebenzise i-multimodal yekhwalithi ephezulu.

Umsebenzi weselula nokuphathwa kwamadokhumenti: thatha izithombe noma ungenise ama-PDF ngokuqondile kuselula noma kuthebhulethi, uqede ngesikhathi sangempela ukubona umbhalo, ukukhipha amathebula, ukuqonda amagrafu nokuphendula imibuzo ngamadokhumenti. Akudingeki ulayishe efwini, okuvikela ubumfihlo futhi kuthuthukise ukusebenza kweselula.
Usizo lwezemfundo olungaxhunyiwe ku-inthanethi: abafundi bangathatha izithombe zemibuzo ukuze bayixazulule, benze ukucabanga ngezibalo futhi babuze imibuzo yesayensi kumadivayisi angenayo inethiwekhi. Ukusebenza kwemodeli ye-35B kudivayisi kwenza umfundisi we-AI wekhwalithi ephezulu afinyelele nasezindaweni ezinezimo zenethiwekhi ezibuthaka.
Umbono wezimoto namarobhothi: ngenxa yokuhambisana nama-chip kusukela ku-Day-0, imodeli ingasetshenziswa ezinhlelweni zezimoto nakumatheminali amarobhothi ukuze ibone indawo, iqonde izithiyo futhi isekele ukuxhumana okubukwayo phakathi komuntu nomshini ngesikhathi sangempela.
Ukuqonda nokuhlola okuqukethwe kwamavidiyo: ikhono lokuqonda amavidiyo amade lisekela ukufingqa okuqukethwe, ukuphendula imibuzo ngamafreyimu abalulekile nokuhlola ukuhambisana kwamavidiyo amade aqoshwe kumadivayisi, lilungele ukuhlolwa kwezokuphepha, ukudala okuqukethwe nezinye izimo.