Umthombo ovulekile
Ithegi yethuluzi le-AIEZZ
Ubheka imikhiqizo ye-AI enethegi ethi “Umthombo ovulekile”, enemiphumela efanayo engu-29.
SmartSub – Ithuluzi ledeskithophu lomthombo ovulekile lokucubungula imibhalo engezansi yomsindo nevidiyo konke endaweni eyodwaUkuqaphela inkulumo nge-AII-SmartSub (妙幕) iyithuluzi ledeskithophu lomthombo ovulekile lokucubungula imibhalo engezansi yomsindo nevidiyo konke endaweni eyodwa. Leli thuluzi lihlanganisa lonke uhlelo lokusebenza lube kuhlelo olulodwa, lisebenzisa amamodeli endawo afana ne-Whisper, FunASR ne-Qwen3-ASR ukuguqula inkulumo ibe umbhalo ngaphandle kwe-inthanethi, lisekela izinsizakalo zokuhumusha ezingaphezu kuka-20 kanye nokuphimisa kwe-AI okunabalingiswa abaningi. Linomlandi wevidiyo eku-inthanethi owakhelwe ngaphakathi, lihambisana nezinkundla ezifana ne-B站 ne-YouTube, lisekela ukusheshiswa kwehadiwe okufana ne-NVIDIA CUDA ne-Apple Core ML, futhi lisebenza endaweni ku-Windows, macOS ne-Linux.00
InstructAV2AV – Imodeli yokuhlela okuhlangene kwevidiyo nomsindo evulwe yi-Zhiyuan ngokubambisana ne-Peking UniversityUkuhlela izithombeI-InstructAV2AV – Imodeli yokuhlela okuhlangene kwevidiyo nomsindo evulwe yi-Zhiyuan ngokubambisana ne-Peking University iyimodeli yokuhlela okuhlangene kwevidiyo nomsindo eyethulwe ngokubambisana yi-Beijing Academy of Artificial Intelligence ne-Peking University. Umsebenzisi udinga kuphela umyalo owodwa wolimi lwemvelo ukuze ahlele ngesikhathi esisodwa izithombe zevidiyo nomsindo ohambisana nazo ngesikhathi sokukhiqiza kusukela ekuqaleni kuye ekugcineni, ngaphandle kokusebenzisa imaski ngesandla noma ukucubungula ngezinyathelo. Imodeli isekela ukuguqulwa kwemigqa yenkulumo, ukushintshwa kwabantu, ukufakwa nokususwa kwezinto, kuyilapho igcina umsindo wangemuva nowemvelo ungashintshi futhi igcina ukuvumelana kwesikhathi phakathi kwesithombe nomsindo. Imisebenzi eyinhloko ihlanganisa ukuguqulwa kwenkulumo kugcinwa ubuwena, ukushintshwa kwezibonelo zevidiyo nomsindo, ukufakwa kwezibonelo, ukususwa kwezibonelo kanye nokuhlelwa okuhlanganisa izici. Ifanele ukuhlelwa kwamafilimu nethelevishini, ukukhiqizwa kwamavidiyo amafushane, ukuphathwa kwabantu ababonakalayo, ukuklama izikhangiso kanye nokukhiqizwa kokuqukethwe okusebenzisanayo.00
JoyAI-Video-EditUkuhlelwa kwamavidiyo nge-AII-JoyAI-Video-Edit – imodeli ka-JD.com yomthombo ovulekile yokuhlela amavidiyo ngokusakaza ngesikhathi sangempela, iyimodeli eyakhiwe futhi yavulwa ngu-JD.com. Isekelwe ekwakhiweni kwe-autoregressive diffusion enamapharamitha angu-16B, ifinyelela isivinini sokucubungula esingu-30FPS ngokulungiswa kwe-720P futhi isekele ukuhlela okusakazwayo okuzinzile kwanoma ibuphi ubude bevidiyo. Abasebenzisi bangasebenzisa imiyalelo yolimi lwemvelo ukuze baguqule abantu, izigcawu, izitayela nezinto ngesikhathi ividiyo idlala, ngaphandle kokulinda ukuthi ividiyo ephelele ikhiqizwe. Ekuhlolweni kwe-OpenVE-Bench, imodeli idlula zonke izindlela zokuhlela ezisakazwayo futhi ihola kuzo zonke izinkomba, kanti iphinde inikeze indlela yokukhiqiza idatha ngezinga elikhulu yobuhlakani obugxilile emzimbeni.00
Orchard – Uhlaka lwe-Agentic AI oluvulelekile olwakhiwe yi-Microsoft ResearchI-ejenti ye-AII-Orchard – uhlaka lwe-Agentic AI oluvulelekile olwakhiwe yi-Microsoft Research. Uhlaka lugxile kusevisi yezindawo i-Orchard Env esekelwe ku-Kubernetes, futhi lusekela ukusetshenziswa kabusha kwama-sandbox ezizindeni ezahlukene ukuze kwenziwe i-data distillation, i-reinforcement learning rollout nokuhlola. Luhlanganisa izimo ezintathu ezinkulu: ubunjiniyela besoftware, ukuzulazula ku-browser kanye nomsizi womuntu siqu. Lunikeza ama-recipe aphelele okuqeqesha e-SFT + RL, kanye nama-trajectory angu-107K e-SWE nedatha engu-3,070 ye-GUI multimodal rollout, futhi lungakwazi ukuqeqeshwa ngokuqondile kusukela ekuqaleni kuye ekugcineni kuma-harness angempela afana ne-Codex ne-OpenClaw. Lubhekiswe emaqenjini ezemfundo nezimboni acwaninga futhi athuthukisa ama-Agent, i-Orchard inezindleko eziphansi zekhompyutha nezokusingathwa, futhi ikhono layo lokusebenza kuma-harness ahlukahlukene lisiza amamodeli amancane ukusondela ekusebenzeni kwezinhlelo ezihamba phambili.00
I-Alpamayo 2 Super – Imodeli ye-AI yokucabanga ngokushayela okuzenzakalelayo evulekile ye-NVIDIAUkuhlela nge-AII-Alpamayo 2 Super iyimodeli ye-AI yokucabanga ngokushayela okuzenzakalelayo evulekile ye-NVIDIA, eyakhiwe phezu kwe-NVIDIA Cosmos 3 Super Reasoner, enamandla okubona indawo ezungezile engu-360°, ukwenza izinqumo ezithuthukile zokushayela nokukhiqiza amalebula okucabanga ngokuzenzakalelayo. Imodeli ikleliswe endaweni yokuqala esilinganisweni sokucabanga ngokushayela okuzenzakalelayo se-LingoQA, futhi isebenzisa ilayisensi evulekile ye-OpenMDW-1.1, esekela ukulungiswa okuqondene nomsebenzi, ukwakhiwa kwemikhiqizo esuselwe kuyo kanye nokusatshalaliswa kabusha kwezentengiso. Iphinde inikeze ngezinguqulo ze-Alpamayo 1.5 / 1 zokuthuthukisa emafini nokusebenzisa imodeli encishisiwe ezimotweni, iqondiswe kuma-Robotaxi nasezimotweni ezishayela ngokuzenzakalelayo.00
wigoloI-ejenti ye-AII-wigolo iyithuluzi lokusesha lendawo elinomthombo ovulekile, elixhuma kuma-Agent okuhlela afana ne-Claude Code ne-Cursor nge-MCP. Ayidingi i-API Key, futhi ngaphandle kwezindleko ungenza usesho, uthwebule amakhasi ewebhu, ukhase isayithi lonke, ukhiphe idatha ehlelekile, ugcine inqolobane futhi wenze ucwaningo endaweni yangakini...00
MemHarness – uhlaka lokwakha kabusha inkumbulo oluqalwe yi-Shanghai AI Lab nabanyeI-ejenti ye-AII-MemHarness iwuhlaka lokwakha kabusha inkumbulo lwama-LLM Agent, olwethulwe ngokuhlanganyela yi-Shanghai AI Lab namanyuvesi ahlanganisa i-Zhejiang University, i-Fudan University kanye ne-Shanghai Jiao Tong University. Ifaka izinyathelo zokuhlaziya nokwakha kabusha phakathi kokusesha inkumbulo nokwenza isenzo, igcine, ibhale kabusha noma ilahle okuhlangenwe nakho kwangaphambilini ngokuhambisana nomongo wamanje. Isebenzisa ukuqeqeshwa kwe-GRPO kusukela ekuqaleni kuye ekugcineni ngaphandle kwedatha eyengeziwe enelebula eyenziwe abantu. Uhlaka lusekela yonke inqubo yokusesha inkumbulo, ukuhlaziya nokwakha kabusha, ukukhiqiza izenzo, ukubuyisela okuhlangenwe nakho nokususa inkumbulo enenani eliphansi, futhi lunikeza inqubo yokuthatha izinqumo enezigaba ezinhlanu. Imodeli enamapharamitha angu-7B idlula i-Gemini-2.5-Pro ngo-23.1% ku-ALFWorld nango-39.7% ku-WebShop, kanti izinga layo lokuphumelela elijwayelekile ezimweni ezingaphandle kokusatshalaliswa lifinyelela ku-85.9%. Isebenziseka emisebenzini yasekhaya yamarobhothi ahlakaniphile, ekuthengeni ku-inthanethi ngokuzenzakalelayo, ekusekelweni kwamakhasimende okuhlakaniphile, ekusizeni ukuthuthukiswa kwekhodi nakwezinye izimo.10
ForgeStencil – uhlelo lwe-Stencil lokucwaninga nokuthunyelwa okuzenzakalelayo ngokuphelele oluqaliswe yi-Mianbi IntelligentI-ejenti ye-AII-ForgeStencil – uhlelo lwe-Stencil lokucwaninga nokuthunyelwa okuzenzakalelayo ngokuphelele oluqaliswe yi-Mianbi Intelligent, lwakhiwe ngokubambisana yi-Mianbi Intelligent nomphakathi we-OpenBMB. Lusebenzisa ukwakheka okuvaliwe kwama-agent amabili, i-Kernel Agent ne-App Agent. Lungenza ngokuzenzakalelayo yonke inqubo kusukela ekutholeni amasu okuthuthukisa, ekuhlanganiseni ama-CUDA Kernel asebenza kahle, kuya ekuhlanganiseni nasekuqinisekiseni izinhlelo ezisezingeni lokukhiqiza, ngaphandle kokungenelela komuntu. Esontweni elilodwa, uhlelo luthuthukise izinhlelo ezingaphezu kwe-100 zokubala zezimboni nezesayensi kusukela ekuqaleni kuye ekugcineni, lwafinyelela ekusheshiseni okumaphakathi okuphindwe ngo-1.41. Luhlanganisa imikhakha yezimboni engu-8 kanye nemikhakha yesayensi engu-5, okuhlanganisa ukuhlola uwoyela negesi, ukulingisa amandla kazibuthe kagesi, nezithombe zezokwelapha, ngaleyo ndlela lwehlise kakhulu ukuncika kochwepheshe be-HPC nesikhathi sokuthuthukisa.00
dots.ttsUkwenziwa kwenkulumo nge-AII-dots.tts iyimodeli eyisisekelo yokwenziwa kwenkulumo ye-autoregressive eqhubekayo ngokuphelele, enamapharamitha ayizigidigidi ezi-2, ekhishwe ngomthombo ovulekile ngokubambisana yiqembu le-Xiaohongshu dots kanye ne-X-LANCE Laboratory yase-Shanghai Jiao Tong University. Imodeli ikhiqiza umsindo we-48kHz ngamaqoqo ngokuqondile endaweni efihlekile eqhubekayo, futhi...00
ShieldstralAI Umhlalelo WezibuzoI-Shieldstral – imodeli yomthombo ovulekile ye-Mistral AI yokuhlukanisa ukuphepha kokuqukethwe kwe-multimodal iyimodeli engu-3B yamapharamitha eyethulwe yi-Mistral AI, eyakhelwe phezu kwe-Ministral-3B. Imodeli ichaza kabusha ukuhlolwa kokuqukethwe kwezigaba ezimisiwe kwendabuko kube umsebenzi wemibuzo nezimpendulo ezimbili, futhi ivumela izinqubomgomo zokuhlola ukuthi zichazwe ngesikhathi sangempela ngemibuzo yolimi lwemvelo ngaphandle kokuqeqeshwa kabusha, ukuze ivumelane nezimo ezahlukene. I-F1 emaphakathi yemodeli ifinyelela ku-84.9% ebhentshimakhi yokuphepha kombhalo, kanti i-F1 yokuphepha kwe-multimodal ifinyelela ku-83.8%, zombili zisezingeni le-SOTA. Ingafakwa kusetshenziswa i-GPU eyodwa engu-16GB futhi isekela izilimi ezingu-12. Imisebenzi eyinhloko ihlanganisa ukuhlola okuvumelana nezinqubomgomo, ukutholwa okuhlanganisiwe kwe-multimodal, ukuhlonza ukwephulwa kwemithetho ngezinga elinemininingwane, amaphuzu okuphepha alinganisiwe, kanye nokufakwa okulula kumadivayisi asemaphethelweni. Ifanele ukulawulwa kobungozi bokuqukethwe ezinkundleni zokuxhumana, ukuvikelwa kokuphepha kwezingxoxo ze-AI, ukuhambisana kwezikhangiso nokumaketha, ukuhlungwa kokuqukethwe kwemfundo eku-inthanethi, kanye nokuhlolwa kokuphepha kwemibhalo yezinkampani.00
PAST-BenchI-ejenti ye-AII-PAST-Bench iwuhlolo lokulinganisa olwakhiwe yithimba likaMengdi Wang eNyuvesi yasePrinceton, olusetshenziselwa ukuhlola ikhono le-Personal AI Agent lokuzithuthukisa ngokuphindaphindiwe. I-PAST-Bench iqhathanisa ukusebenza kwemisebenzi ngaphansi kwezimo ezinememori nezingenayo, ukuze inqume ukuthi ulwazi olugcinwe yi-Agent phakathi kwamaseshini luthuthukisa ngempela yini ukuziphatha okulandelayo...00
PixelRAGInkundla yokuthuthukisaI-PixelRAG iwuhlaka lwe-RAG olugxile ekuboneni, olukhishwe njenge-open source yi-SkyLab/BAIR yase-Berkeley. Esikhundleni sendlela evamile yokukhipha amakhasi ewebhu abe umbhalo omsulwa bese kuwasesha, uhlaka lusebenzisa isiphequluli ukuguqula amakhasi ewebhu nama-PDF abe amathayela ezithombe-skrini, lusebenzisa i-Qwen3-VL ecushwe nge-LoRA...00
LTX-2.5Ukukhiqiza ividiyo nge-AII-LTX-2.5 iyimodeli eyisisekelo ye-AI yokukhiqiza amavidiyo evulekile ye-LTX, enamapharamitha angu-22B futhi enesisindo esivulekile kusukela ngosuku lokukhishwa. Imodeli isekela ukukhiqizwa kwamashothi amaningi ngokwemvelo, i-4K HDR kanye nezindlela zokusebenza zobuchwepheshe ze-RAW/EXR, futhi ingakhiqiza umsindo ngesikhathi esifanayo, ...00
HOMIEUkukhiqiza ividiyo nge-AII-HOMIE iwuhlaka oluvulekile lokukhiqiza amavidiyo abantu bedijithali olukhishwe yi-Hong Kong University of Science and Technology, olwakhelwe kunethiwekhi eyinhloko i-Wan2.1-T2V-14B futhi oluhlanganisa imodeli enkulu yezindlela eziningi i-Qwen3-VL. Uhlaka luphatha ngokuhlangene izici ezine: umuntu wedijithali + umkhiqizo + i-Logo + umbhalo, ...00
Palmier ProUkuhlelwa kwamavidiyo nge-AII-Palmier Pro iwumhleli wamavidiyo womdabu we-macOS owakhelwe inkathi ye-AI, wakhiwe kusukela ekuqaleni nge-Swift, ungumthombo ovulekile futhi izici eziyisisekelo zimahhala.00
MindMemOS – Uhlelo lokusebenza lwememori lwe-AI Agent oluvulekele umphakathi, oluvulwe yi-Huawei Noah's Ark LabI-ejenti ye-AIMindMemOS – Uhlelo lokusebenza lwememori lwe-AI Agent oluvulekele umphakathi, oluvulwe yi-Huawei Noah's Ark Lab, lwenzelwe ukuhlinzeka ama-agent adinga ukuphathwa kwememori yesikhathi eside ngongqimba lwememori oluvulelekile nolungasetshenziswa kabusha. Lusebenzisa isakhiwo esinezilinganiso ezintathu: ibhizinisi, isibaluli nesikhathi, luhlukanisa imemori ku-Agent eyodwa luyenze impahla ezimele esebenza kuwo wonke amafreyimwokhi, futhi luthuthukisa imemori ngokuqhubekayo ngezinqubo ezifana ne-Dreaming yokuhlela ungaxhunyiwe ku-inthanethi, i-Feedback yokulungisa amaphutha kanye ne-Skill Evolution yokuziphendukela ngokuzenzakalelayo. Uhlelo luhlinzeka ngezindlela ezimbili zokubhala imemori (MindVanilla, MindSchema) kanye nokusesha kwe-Compact Search okunemizila eminingi, futhi lusekela ukufakwa ku-Docker endaweni, i-API esebenza efwini, i-Python SDK, i-CLI nezinye izindlela zokuxhuma. Ngokunemba kokubuyisa okuhlelekile okungaphezu kuka-94% kanye nokuthuthukisa ukunemba kwezimpendulo zemibuzo ngo-10% ngemva kokucindezelwa kwememori, lusiza abathuthukisi ukuthi babelane ngolwazi phakathi kwezingxoxo futhi bathuthukise ukusebenza kwemisebenzi ezimweni ezifana nabasizi abenziwe ngezifiso, ukubambisana kwama-Agent amaningi kanye ne-automation eyinkimbinkimbi yemisebenzi yasehhovisi.00
Nemotron 3.5 LightningI-ejenti ye-AII-Nemotron 3.5 Lightning iyimodeli ye-MoE yomthombo ovulekile enamapharamitha angu-30B eyethulwe ngu-NVIDIA, elungiselelwe ngokukhethekile amasistimu anama-agent amaningi. Uma iqhathaniswa namamodeli afanayo, ijubane lokukhipha imiphumela likhuphuka izikhathi ezi-4, kanti ijubane lokuqedela imisebenzi yama-agent likhuphuka ngo-30%, ...00
Wan-Animate-2 – imodeli yesizukulwane esisha yokugqwayiza abalingiswa ekhishwe njenge-open source yiqembu le-WanxiangUkwenza opopayiI-Wan-Animate-2 iyimodeli yesizukulwane esisha yokugqwayiza abalingiswa ekhishwe njenge-open source yiqembu le-Wanxiang. Njengokuthuthukiswa okukhulu kwesakhiwo se-Wan-Animate, isebenzisa i-DiT enamagatsha amabili esebenza ekugcineni kuya ekugcineni, ngaphandle kwesidingo sohlaka lokuma olucacile, ukuze ithwebule ulwazi lokunyakaza oluvela kuvidiyo eyisithenjwa. Isekela ukuguqula ividiyo eqondisayo ibe ukugqwayiza komlingiswa, ukugcina ubuwena nokubukeka, kanye nokulawula i-engeli yokubuka okuqhutshwa umbhalo. Inikeza izindlela ezimbili zokwenza inference: i-Base yekhwalithi ephezulu kanye ne-Distillation enezinyathelo ezimbalwa, futhi yethula uhlobo lwe-Lite oluqondiswe ekugqwayizeni abalingiswa okusakazwa ngesikhathi sangempela. Le modeli ifanele izimo ezifana nabasakazi ababonakalayo, ukwakhiwa kwamavidiyo amafushane, nabantu bedijithali, ihlose ukuthuthukisa kokubili ukwethembeka kokunyakaza nokuvumelana kobuwena, kuyilapho yehlisa ukubambezeleka kokukhiqiza.00
SenseNova U1.5-Lite-Preview – Imodeli ye-multimodal elula, evulekile yakwa-SenseTimeAmamodeli amakhulu e-AISenseNova U1.5-Lite-Preview – imodeli ye-multimodal elula, evulekile yakwa-SenseTime, isekelwe ekwakhiweni kwe-NEO-Unify futhi isebenzisa amapharamitha angu-8B-MoT ukuze ihlanganise ukuqonda okubukwayo, ukucabanga, ukwakha nokuhlela ngendlela yemvelo. Imodeli isekela ukwakha izithombe ze-ultra-HD ze-4K ngendlela yemvelo, inikeza ukuthungwa okunemininingwane, ukubukeka okungokoqobo kanye nekhono lokunikeza umbhalo wesiShayina nesiNgisi ezakhiweni eziyinkimbinkimbi. I-Prompt Enhance Skill eyakhelwe ngaphakathi yehlisa umngcele wokudala, ikwazi ukusebenzisa ngokunembile imiyalelo emide kakhulu yolimi lwemvelo kanye nemiyalelo ebonakalayo ehlelekile, futhi isekele ukuhlela okuqhubekayo nokuphindaphindwayo. Iqondiswe kubaklami, abathuthukisi nabadali bokuqukethwe, inikeza ithuluzi lokudala okubukwayo elisezingeni lezentengiso, elinezindleko eziphansi zokufakwa kanye nezinga eliphezulu lokulandela imiyalelo.10
IndexTTS-2.5Ukwenziwa kwenkulumo nge-AII-IndexTTS-2.5 iyimodeli yokukopisha izwi esezingeni lezimboni ye-zero-shot, ekhishwe yi-Bilibili njengomthombo ovulekile. Inamapharamitha angu-0.8B kuphela, futhi isekela ukudluliselwa kwezwi phakathi kwezilimi ezinhlanu: isiShayina, isiNgisi, isiJapane, iSpanishi nesi-Arabhu.00
MAGI-2-preview – Imodeli yokukhiqiza amavidiyo e-multimodal evulekile ye-Sand.aiUkukhiqiza ividiyo nge-AIMAGI-2-preview – Imodeli yokukhiqiza amavidiyo e-multimodal evulekile ye-Sand.ai iyimodeli yokuqala emhlabeni yokukhiqiza amavidiyo e-multimodal ye-MoE enamapharamitha ayizigidigidi eziyikhulu, evulwe yi-Sand.ai, enamapharamitha aphelele angu-114B kodwa evula angu-6B kuphela. Imodeli iqhubeka nokusebenzisa isakhiwo se-single-stream, ihlanganisa umbhalo, ividiyo nomsindo ku-Transformer eyodwa ukuze ikwenze ukumodela okuhlangene, ngaleyo ndlela ikhiqize i-multimodal ngokwemvelo. Imisebenzi yayo esemqoka ihlanganisa ukukhiqiza kusuka embhalweni, esithombeni, nasevidiyweni kuya kwenye ividiyo, kanye nokuqonda umsindo okuhlanganiswe ngokwemvelo. I-MoE sparse activation yehlisa kakhulu izindleko zokwenza inference. Abasebenzisi abaqondisiwe ngabacwaningi be-AI, abathuthukisi namaqembu ezobuchwepheshe ezinkampani. Isebenza ocwaningweni lwezemfundo, ekusetshenzisweni kwangasese nasekuhlolweni kwangaphambili kwemikhiqizo yamafilimu nezinye izimo ze-multimodal, inikeze isixazululo esisebenza kahle sokukhiqiza amavidiyo esivulekile futhi esingasetshenziswa kwezentengiso.00
Qwen-CUA – I-Computer Use Agent yomdabu eyethulwe yi-Alibaba Qwen nabanyeI-ejenti ye-AII-Qwen-CUA – I-Computer Use Agent yomdabu eyethulwe yi-Alibaba Qwen nabanye, isekelwe ekwakhiweni kwe-Mixture-of-Experts engu-397B-A17B. Ibona isimo sesixhumi esibonakalayo ngezithombe zesikrini kuphela, ngaphandle kokuncika esihlahleni se-DOM noma kumethadatha yokufinyeleleka, futhi ikhiphe ngokuqondile imicimbi yekhibhodi negundane ukuze ilawule iziphequluli, amadeskithophu nezinhlelo zobuchwepheshe. Ekuhlolweni kwe-OSWorld-Verified ithole amaphuzu angu-86.2, kanti inguqulo enkulu i-Qwen-CUA-Max ithole angu-87.6, okuyizinga elihamba phambili emkhakheni wama-CUA omthombo ovulekile. Isebenzisa ukuqeqeshwa kokuphela-kuya-ekupheleni kokubona-isenzo, inokusebenziseka ezinkundleni ezahlukene nokuzinza okuphezulu, futhi sekukhishwe umbiko wezobuchwepheshe kanye ne-Demo, okuyenza ifanele ukuhlolwa okuzenzakalelayo, i-RPA, usizo lokufinyeleleka kanye nokulawula izinhlelo zocwaningo.00
SALMONN-2 – imodeli yolimi olukhulu yomsindo ejwayelekile ekhishwe yi-Tsinghua nabanyeUkuhlela nge-AII-SALMONN-2 – imodeli yolimi olukhulu yomsindo ejwayelekile ekhishwe yi-Tsinghua nabanye iyimodeli yolimi olukhulu yomsindo ejwayelekile eyakhiwe kusetshenziswa i-SPEAR unified self-supervised audio encoder yakwa-ByteDance kanye ne-multi-layer feature fusion adapter, ne-Qwen3 njengesisekelo sombhalo. Isekela imisebenzi ehlukahlukene efana nokuqaphela inkulumo, ukuchaza umsindo, ukuqonda umculo, ukuqaphela imizwa nokuqinisekisa isikhulumi, futhi ngokokuqala ngqa ku-ALLM ejwayelekile ihlanganisa ngokuhlelekile amakhono athuthukile okuhlonza izehlakalo zomsindo, ukuthola umsindo okhohlisiwe, ukuhlola ikhwalithi yenkulumo kanye nokuqaphela inkulumo okunomongo we-multimodal. Imodeli isebenzisa i-SPEAR encoder eyodwa esikhundleni sohlelo oluvamile lwama-encoder amabili, futhi ihlanganisa izici zomsindo zamazinga amaningi nge-MLF adapter ukuze ithole ukusebenza okulinganiselayo ezizindeni ezahlukene. Ngokuqeqeshwa okuphumelelayo kusetshenziswa cishe amahora angu-18,000 edatha egadiwe kuphela, i-SALMONN-2 ifinyelela ukusebenza okuhamba phambili phakathi kwamamodeli omthombo ovulekile anosayizi ofanayo kuma-benchmark amathathu aphelele i-MMAU-Pro, i-MMAR kanye ne-MMSU. Ilungele izimo ezifana nabasizi bemihlangano abahlakaniphile, ukuhlolwa kokuqukethwe komsindo kanye nokuqapha ikhwalithi yenkulumo.00
Muse Glimmer – Imodeli yolimi enkulu enamapharamitha ayizigidi eziyizinkulungwane ezingu-30 evulwe yi-MetaI-ejenti ye-AIIMuse Glimmer imodeli yolimi enkulu enamapharamitha ayizigidi eziyizinkulungwane ezingu-30 evulwe yi-Meta, eyenzelwe futhi yalungiselelwa ngokujulile ukuqhuba ama-Workflow e-Agent ahlala kudivayisi ubusuku nemini. Ngobuchwepheshe be-4bit quantization, imodeli ingasebenza kahle ku-GPU eyikhadi elilodwa elinememori engu-24GB noma ku-Apple Silicon (M4/M5 Max) enememori ehlanganisiwe engu-32GB, cishe ngaphandle kokuncipha kokusebenza. Ihlanganisa i-DFlash speculative decoding, efinyelela ekusheshiseni okufika ku-3.1x ku-hardware esezingeni eliphezulu njenge-RTX 5090, futhi isekela imisebenzi ye-multimodal efana nezingxoxo zesikhathi sangempela, ukubiza amathuluzi, ukukhiqiza ikhodi kanye nokucabanga ngezinyathelo eziningi. Ifanele abathuthukisi ngabanye, amabhizinisi amancane naphakathi, kanye nezimboni ezifana nezempilo nezezimali ezinezidingo eziphezulu zobumfihlo bedatha. Inikeza amandla e-AI asebenza kahle endaweni ngaphandle kwemfuneko enkulu yememori ye-GPU, futhi yonke idatha ihlala ngaphakathi kwedivayisi.00
Hunyuan3D-Buffalo 1.0 – Uhlaka lwe-3D lwezindlela eziningi olukhishwe yi-Tencent HunyuanUmklamo we-3DI-Hunyuan3D-Buffalo 1.0 – Uhlaka lwe-3D lwezindlela eziningi olukhishwe yi-Tencent Hunyuan iwuhlaka oluhlanganisiwe lwe-3D lwezindlela eziningi olukhishwe yi-Tencent Hunyuan. Ihlanganisa imibuzo nezimpendulo ze-3D, ukubeka indawo, umbhalo uye ku-3D, ukuhlela okuqondiswa yimiyalelo kanye nokukhiqizwa kwezingxenye kube yinqubo eyodwa isebenzisa i-Hunyuan3D-VLM backbone eyabiwe. Uhlaka lusekela ukuqonda isakhiwo samamodeli e-3D ngolimi lwemvelo, ukuthola izingxenye, ukuguqula ijiyomethri yendawo ngokwemiyalelo, kanye nokukhipha izingxenye ezisezingeni lencazelo. Luhlose ukuhlanganisa ukuqonda, ukukhiqiza nokuhlela kwe-3D kube umjikelezo ophelele, lunikeze umkhiqizo wempahla ye-3D ohlanganisekayo wemidlalo, opopayi nokuklanywa kwezimboni. Imisebenzi esemqoka ihlanganisa ukuqonda kwe-3D, umbhalo uye ku-3D, ukuhlela okuqondiswa yimiyalelo kanye nokukhiqizwa kwezingxenye; zonke lezi zinhlobo ezine zabelana ngomehluko ofanayo wencazelo ye-3D kanye nendlela yokucubungula. Lufanele abathuthukisi bemidlalo, abakhiqizi bopopayi, abaklami bezimboni nabadali bokuqukethwe kwe-3D. Ngokusebenzisana ngolimi lwemvelo, lwehlisa izithiyo zobuchwepheshe futhi luthuthukise ukusetshenziswa kabusha kwempahla kanye nokusebenza kahle kokuphindaphinda.00
ShotcutUkuhlelwa kwamavidiyo nge-AII-Shotcut iyisofthiwe evulekile yokuhlela amavidiyo esekelwe ku-FFmpeg, imahhala ngokuphelele futhi ayinawo ama-watermark noma imikhawulo yobude besikhathi. Isekela ukuhlela ngokuqondile ngaphandle kokungenisa amakhulu amafomethi omsindo namavidiyo, inikeza umugqa wesikhathi onamathrekhi amaningi, izithombe ezinyakazayo ezilawulwa ngama-keyframe, izinhlobo eziningi...00
WorldClaw – uhlaka lokukhiqiza imihlaba ye-3D olwenziwe ithimba le-Tencent HunyuanUmklamo we-3DI-WorldClaw iwuhlaka lokukhiqiza imihlaba evulekile ye-3D oluqhutshwa yi-Agent, olwenziwe ithimba le-Tencent Hunyuan. Lapho umsebenzisi efaka incazelo ngolimi lwemvelo, uhlelo lungahlela ngokuzenzakalelayo indawo, izifunda, izakhiwo, izinto kanye nobudlelwano bendawo, lusebenzisa isu lokukhiqiza elisuka kokujwayelekile liye kokuningiliziwe, bese lwakha isigcawu esiphelele se-3D ku-Blender esingahlolwa ngokukhululekile futhi esisekela ukuhlelwa kwento ngayinye. Uhlaka luqhutshwa ngokubambisana ngamamodeli amaningi afana ne-Claude Opus, GPT-Image-2, SAM3 kanye ne-Hunyuan3D, futhi lunekhono le-visual Agent lokuzihlola nokuzilungisa, olungalungisa ngokuzenzakalelayo izinkinga zokuthintana ezifana nezinto ezilenga emoyeni noma ezingena kwezinye.00
PyTorchAmathuluzi Okuhlela nge-AII-PyTorch iwuhlaka oluvulekile lokufunda komshini olwasungulwa ithimba locwaningo le-Meta AI, futhi manje luphethwe yi-Linux Foundation. Lusebenza kakhulu kubacwaningi bezobuhlakani bokwenziwa, onjiniyela bokufunda komshini, ososayensi bedatha nabafundi. Abasebenzisi bavame ukusebenzisa i-Python ukuchaza amamodeli nezinqubo zokubala eziya phambili, bese besebenzisa ukubala kwama-tensor, i-automatic differentiation namamojuli amanethiwekhi ezinzwa ukuze baqeqeshe, kanye nemitapo yolwazi ehlobene ukuphatha imisebenzi yombono, umsindo noma umbhalo. Igrafu yokubala eyakhiwa ngesikhathi sokusebenza yenza ukuhlola nokulungisa kube lula, futhi isekela ukuqhubekisela phambili i-prototype yocwaningo iye ekusetshenzisweni kwemodeli. Ukusetshenziswa kwangempela kusadinga ulwazi lwe-Python, ukubala kwama-tensor nezisekelo zokufunda okujulile; lapho ikhodi yokuhlola idluliselwa endaweni yokukhiqiza, kudingeka futhi uphathe ngokwakho ukuthunyelwa kobunjiniyela nokumiswa kwezinsiza.057.6
Qwen 3.8-Max – imodeli yolimi ehamba phambili eyethulwe yiQwen Team yakwa-AliyunAmamodeli amakhulu e-AII-Qwen 3.8-Max – imodeli yolimi ehamba phambili eyethulwe yiQwen Team yakwa-Aliyun, eyakhiwe ngokwandisa i-architecture ye-Qwen 3.5 yafinyelela kumapharamitha angu-2.4 trillion, futhi ihloselwe ukuba imodeli yolimi esezingeni elikhulu kakhulu yamabhizinisi nocwaningo. Isekela ukubhala ikhodi okuzenzakalelayo isikhathi eside, ukuphinda izivivinyo zamaphepha ocwaningo, imincintiswano, ukuthuthukisa ukusebenza kwehhovisi, ukuthuthukiswa kwamasu e-quantitative kanye nokuziklanyela kwama-chip. Imodeli isebenzisa i-reinforcement learning yomhlaba wangempela kanye ne-feedback loop yokuzithuthukisa ukuze ikwazi ukusebenza ngokuzenzakalelayo izinsuku ezimbalwa kuya emashumini ezinsuku, iqede imisebenzi eyinkimbinkimbi ngaphandle kokungenelela komuntu. Abasebenzisi bangayibiza nge-API ehambisana ne-OpenAI noma i-Anthropic protocol esesiteji se-Qianwen AI, noma basebenzise ama-open-source weights endaweni, ukuze bayihlanganise kalula nama-agent frameworks afana ne-Claude Code, Codex kanye ne-Qoder. Ngenxa yalokho, amaqembu ocwaningo nentuthuko, ochwepheshe bezomthetho nokuhambisana nemithetho, abaklami nabanye abasebenzisi abangochwepheshe bangakhuphula kakhulu ukusebenza kwabo, banciphise amaphrojekthi athatha isikhathi eside abe ukulethwa ngaphakathi kwengxoxo eyodwa.10Wonke amathuluzi analeli thegi ayaboniswa.