I-Tencent Hy4: Imodeli ye-Open-Weight ye-770B Nekutsi Isho Kutsini Kubhizinisi Lakho

Imphendvulo lecondzile: mhla ti-28 Agasti 2026 i-Tencent yashicilela i-Hy4 preview njengemodeli ye-open-weight ngaphansi kwelayisensi ye-Apache 2.0, lenema-parameter langu-770 billion sewuphelele, lapho langu-49 billion kuphela asebenta nge-token ngayinye, kanye nefasitelo le-context lelifika kuma-token lasigidi sinye. Ngekwemkhuba, loko kusho kutsi nobe nguyiphi inkampani ingayilandzisa imodeli futsi iyisebentise engcalasizindeni yayo ngaphandle kwemvume yemtsengisi futsi ngaphandle kwemali yelucingo ngalunye. Akusho kutsi ungayisebentisa ku-server lesehhovisi lakho. Ligebe emkhatsini walemisho lemibili linguyo lelincuma kutsi lendzaba iyakutsintsa yini nobe cha.

Lokwamenyetelwa ngempela

Ngaphambi kwanobe nguluphi luhlatiyo, kuyasita kuhlukanisa emaciniso lashicilelwe ekuchazeni. Loko i-Tencent lekukhiphile endzaweni lesemtsetfweni yekugcina imodeli kufaka ekhatsi:

  • Bukhulu nekusebenta: ema-parameter langu-770 billion sewuphelele, langu-49 billion asebenta nge-token ngayinye, ekwakhiweni kwe-mixture-of-experts.
  • Kwakheka kwangekhatsi: tendlalelo letingu-78, sekucala lesiminyene kantsi letinye ti-mixture-of-experts letinetingcweti letingu-256 letitfunyelwako kanye nengcweti yinye lehlanganyelwako, lapho kukhetfwa tingcweti letisiphohlongo nge-token ngayinye.
  • Lifasitelo le-context: ema-token lasigidi sinye, nebukhulu belulwimi lwemagama labu-120,832.
  • Ilayisensi: i-Apache 2.0, ilayisensi levulekile levumela kusetjentiswa kwebhizinisi, kuguculwa kanye nekusabalaliswa kabusha.
  • Luhlobo lolu-quantised: luhlobo lwe-FP8 luyatfolakala eceleni kwemodeli lenekunemba lokuphelele.
  • Kusheshisa i-inference: sendlalelo lesakhelwe ekhatsi sekubikezela ema-token lamanyenti lesisekela i-speculative decoding.
  • Lokugcizelelwe lokushiwo: imodeli ihlose kubhala ikhodi, kukhicita emsebentini wasehhovisi, kutfutfukisa imidlalo kanye nelucwaningo lwesayensi kunekuba yingcoco yonkhe.

Loko lokungazange kumenyetelwe, futsi lokungafanele kucatjangwe, kunobe nguyiphi intsembiso yekuphakama kunema-modeli lavaliwe esimeni sakho lesitsite. Kucatsaniswa lokushicilelwe kuma-benchmark emphakatsi akukutjeli lutfo lolwetsembekile mayelana nekusebenta ngemibhalo yakho, lulwimi lwemakhasimende akho kanye nesimo semkhakha wakho. Leyo mphendvulo itsatfwa kuphela ekuhloleni kwangekhatsi kudatha yakho.

Kungani ilayisensi ye-Apache 2.0 ingasiyo imininingwane lencane yemtsetfo

Inkhulumo lenkhulu mayelana nema-modeli lavulekile igcile ekhwalithini yemiphumela, kantsi inzuzo yangempela kumnikati webhizinisi isesigabeni selayisensi. Ilayisensi levulekile isho tintfo letintsatfu letisebentako: kusebentisa imodeli ngekhatsi kwengcalasizinda yakho akudzingi kuvunywa ngumuntfu, kuyigucula nekuyilungisa ngedatha yakho kuvumelekile, futsi kuchubeka kwemsebenti wakho akuncikanga ekutseni umtsengisi uchubeke ayinikela nobe agcine emanani akhe angashintji.

Leliphuzu lekugcina lingumongo walendzaba. Inkampani leyakhe yonkhe insita yemakhasimende ayo ku-API yinye itfwele ingoti lethulile: kushintja kwemanani, luhlobo lolumisiwe nobe kushintja kwemgomo wekusebentisa kungayibuyisela ekucaleni. Imodeli ye-open-weight ayiyisusi lengoti, kodvwa ikunika indlela yekuphuma: ikhophi loyilandzile ihlala ingeyakho.

Tibalo te-hardware tematfoko langetulu letivame kweciwa

Lapha ingcondvo lechumile iba tinombolo. Ema-parameter langu-770 billion ngekunemba kwe-FP8 asho cishe i-byte yinye nge-parameter ngayinye, lokungema-gigabyte langetulu kwa-700 ema-weight emodeli odvwa ngaphambi kwanobe yini lenye. Engeta inkhumbulo ye-context lekhula ngebude bengcoco, futsi lifasitelo lema-token lasigidi sinye lisho inkhumbulo lengetiwe lenkhulu ekusetjentisweni kwangempela; bese wengeta kuphindzaphindza kute umsebenti ungawi nangabe i-node yinye yehluleka.

Siphetfo kutsi lena yimodeli yesikhungo sedatha, hhayi yemodeli ye-server yasehhovisi. Kutiphatsela i-hosting yayo kusincumo lesidzinga i-hardware lekhetsekile kanye nelicembu lekusebenta lelati kuhlukanisa imodeli kuma-GPU lamanyenti nekulandzelela kubambeka nekusetjentiswa. Leyo yindleko lengashintji lengancipheki nangabe kusetjentiswa kwakho kwehla, ngalokungafani ne-API lapho ukhokhela loko lokusebentisako.

Umtsetfo lowusebentako: kutiphatsela i-hosting kunengcondvo kuphela ngaphansi kwesimo sinye kuletimo letimbili. Kungaba unekusetjentiswa lokukhulu nalokusimeme lokwenta indleko lengashintji ishibhile kunekusetjentiswa lokushintjashintjako, nobe ubhekene nesidzingo semtsetfo nobe sesivumelwano lesivimbela idatha kutsi iphume emngceleni lochaziwe. Ngaphandle kwaleto timo letimbili, kutiphatsela i-hosting kuvame kuba sincumo semiva kunekuba sincumo semnotfo.

Lokwenta kwakheka kube nekuhlaba umxhwele kwetheknoloji

Inombolo lehlaba umxhwele ku-Hy4 akusiyo i-770 billion kodvwa yi-49 billion. Kwakhiwa kwe-mixture-of-experts kusho kutsi imodeli inemandla lamakhulu kodvwa isebentisa kuphela incenye lencane yayo nge-token ngayinye, ngako utfola ikhwalithi yelizinga lelisetulu ngendleko yekubala lesedvute kwemodeli yelizinga lelimaphakatsi. Lendlela yekwakha nguyo leyasondzeza ema-modeli lavulekile kulavaliwe eminyakeni lemibili leyengcile.

Engeta kuloko indlela ye-sparse attention lephindze isebentise ema-index e-sparsity etendlalelweni tonkhe, kanye nesendlalelo lesakhelwe ekhatsi sekubikezela ema-token lamanyenti lesenta kutsi i-speculative decoding ikhoneke kute kufinyetwe sikhatsi sekukhicita. Ngelulwimi lwebhizinisi: lemininingwane yehlisa indleko nge-token ngayinye kanye nekubambeka, futsi leto tintfo letimbili tincuma kutsi sici singasakatwa yini kubasebentisi labatinkhulungwane nobe sihlale singumlinga wangekhatsi.

Tindlela letintsatfu letisebentako kubhizinisi lakho

  • I-API lephatsiwe kumodeli levaliwe: indlela lesheshako kakhulu yekwetfula, ayinamtfwalo wekusebenta, futsi inendleko leshintjashintjako lekhula ngekusetjentiswa. Indlela lejwayelekile etinyangeni tekucala letisitfupha tanobe nguyiphi imkhicito.
  • I-Hy4 ngemhlinzeki lokugcinela yona: utfola kuguculeka kwelayisensi kanye nekhono lekuhamba emkhatsini wabahlinzeki ngaphandle kwekutsenga i-hardware. Lena yindlela lesemkhatsini lesebentako bantfu labanyenti labayishiyako.
  • Kutiphatsela i-hosting ngalokuphelele: kulawula lokuphakeme kakhulu nendleko lephakeme kakhulu. Kuyafaneleka nangabe ubhekene nesidzingo lesicinile sekuhlala kwedatha elive nobe ugijimisa linani lekusebenta lelikhulu nalelisimeme.

Imibuto lencuma sincumo

Ngaphambi kwekusebentisa ngisho i-riyal yinye kulendlela, phendvula imibuto lemine ngekwetsembeka. Nguluphi luhlobo lwedatha loludlula emodelini, futsi ukhona yini sibopho semtsetfo nobe sesivumelwano lesiyivimbela kutsi iphume? Lithini linani lenyanga lelilindzelekile etinyangeni letisitfupha letitako, hhayi evikini lekucala? Wena, nobe umlingani wakho wetheknoloji, ninalo yini likhono langempela lekusebenta emahora onkhe? Futsi uyakhile yini sendlalelo se-abstraction kuhlelo lwakho lokusebenta lesikuvumela kushintja ema-modeli ngaphandle kwekubhala kabusha umkhicito?

Nangabe imphendvulo yembuto wekugcina itsi cha, loko kungukubaluleka kwakho kwekucala ngaphambi kwanobe nguyiphi inkhulumo mayelana nekutsi nguyiphi imodeli lesetjentiswako. Umehluko emkhatsini wenkampani levumelanako naleboshiwe akusikho kukhetsa imodeli lefanele, kodvwa kwakha luhlelo lolugcina lukhetfo lungashintjwa ngemuva ngendleko lefanele. Buka sihloko setfu mayelana nesifundza se-cloud sasekhaya nekuhlala kwedatha elive nangabe kuhlala kwedatha kusizatfu sakho lesikhulu.

Indlela lesikuphatsa ngayo loku e-Origami

Sakha tinhlelo tekusebenta temakhasimende kute imodeli ibe yincenye lengashintjwa ngemuva kwesixhumi sangekhatsi lesifanako, futsi sigcina lisetho lekuhlola lelinemacala langempela emakhasimende lesiwagijimisa kunobe nguyiphi imodeli lelulekelwe ngaphambi kwekuyamukela. Loko kuguca tindzaba letifana nekukhishwa kwe-Hy4 kusuka esihlokweni setheknoloji kube sincumo lesilinganisekako: gijimisa lisetho lekuhlola, ucatsanise ikhwalithi, indleko nekubambeka, bese uncuma ngetinombolo kunekutsi uncume ngemibono.

Imitfombo