Shannon 3.1
Ukuzungeza kokucabanga okufanayo, kususelwe kuqoqo lwethu lwama-GPU: ihlakaniphe ngo-32%, isheshe ngo-10-15x, inewindi lomongo lamathokheni angu-196,608.
Ngamafuphi
I-Shannon 3.1 igcina konke okwenza i-Shannon 3 ifaneleke ukusetshenziswa — ukuzungeza kokucabanga okuphindaphindayo, kungekho ungqimba lokwenqaba, kungekho ukuhlungwa kokukhishwayo — bese ishintsha ukuthi isebenza kuphi nokuthi isebenza kanjani. Imodeli manje ikhonza isuka kuqoqo lwethu lwama-GPU esikhundleni somsingathi wangaphandle wokudekhoda. Ukuhlolwa kwe-Shannon Lab kukala ukuthuthuka kobuhlakani ngo-32% kanye nezimpendulo ezisheshayo ngo-10-15x uma kuqhathaniswa ne-Shannon 3.0. Iwindi lomongo lisuka kumathokheni angu-32,768 laya kwangu-196,608. Ungqimba lokulinganisa isivinini olwaludonsa ngamabomu okukhishwa yi-3.0 selususiwe: i-3.1 isakaza ngesivinini esigcwele senjini. Ama-id emodeli angu-shannon-3.1 no-shannon-3.1-pro, engxoxweni nakuzo zontathu izinhlobo ze-API.
Ukwethulwa kwamamodeli amaningi kukucela ukuba ukholwe yisimangalo samandla bese ulinda ithebula lezilinganiso ukuze lixazulule impikiswano. Lokhu kulula ukukuhlola: vula amathebhu amabili, faka umyalo ofanayo ku-shannon-3 naku-shannon-3.1, bese ubuka. Umehluko wokuthi impendulo ifika ngokushesha kangakanani awufihlekile, futhi akulona iqhinga lokubonisa. I-Shannon 3.0 yayibanjwe ngamabomu. I-Shannon 3.1 ayibanjwe.
01Yini eyashintsha ngempela ku-Shannon 3.1
I-Shannon 3.0 yethula into echaza lo mndeni: ukuzungeza kokucabanga okuphindaphindayo. Imodeli ayiphenduli ngomcabango wayo wokuqala. Iyacabanga, ibhale uhlaka, ibuyekeze uhlaka lwayo ngokumelene nombuzo, bese iluthuthukisa. I-Lite isebenzisa ukudlula okukodwa kwalolo zungeza; i-Pro isebenzisa ukuzungeza okuphelele, kufaka phakathi isinyathelo sokuvuna ulwazi ngaphambi kokubhala uhlaka. Lolo klamo luchazwe ngokuningiliziwe esihlokweni socwaningo se-Shannon 3, futhi alukho kulo olushintshile ku-3.1.
Okushintshile ngumshini ongaphansi. I-Shannon 3.0 yayikhonzwa ngomsingathi wangaphandle wokudekhoda — indlela enengqondo yokwethula imodeli, kodwa indlela evimbelayo yokuyisebenzisa. Uthola ukumisa kwenkonzo komunye umuntu, ulayini komunye umuntu, uphahla lomongo lomunye umuntu, kanye nomqondo womunye umuntu wokuthi uvunyelwe amathokheni amangaki ngomzuzwana. I-Shannon 3.1 isebenza kuqoqo lwethu lwama-GPU, kusitaki senkonzo esilungiswa yithi, futhi zonke izinombolo eziyinhloko kulesi sihloko zivela kulesi sinqumo esisodwa.
| Shannon 3.0 | Shannon 3.1 | |
|---|---|---|
| Isebenza kuphi | Umsingathi wangaphandle | Iqoqo lethu lwama-GPU |
| Iwindi lomongo | 32,768 | 196,608 |
| Ukusakazwa kokukhishwayo | Kulinganisiwe / kunciphisiwe | Isivinini esigcwele senjini |
| Izisindo | Okuzenzakalelayo komsingathi | I-NVFP4 yamabhithi ama-4 |
| Ukudekhoda | Okujwayelekile | Okwe-speculative |
| Ukuzungeza kokucabanga | Cabanga → bhala uhlaka → buyekeza → thuthukisa | Akushintshile |
| Ungqimba lokwenqaba | Alukho | Alukho |
| Ama-id emodeli | shannon-3, shannon-3-pro | shannon-3.1, shannon-3.1-pro |
02Kungani i-Shannon 3.1 izwakala isheshe ngo-10-15x
Inombolo yesivinini yiyona abantu abaqala ngayo ukubuza, ngakho kufanele sichaze ngokucacile ukuthi ivelaphi. Kunemithelela emibili ezimele, futhi omkhulu kuyo yiwona ongakhangi kakhulu.
Sasusa isango lesivinini
Okwakukhishwa yi-Shannon 3.0 kwakudlula kungqimba lokulinganisa isivinini. Amathokheni ayephuma enjinini, abanjwe, bese ekhululelwa kuxhumano lwakho ngokohlelo lwesikhathi. Lokhu kwakungelona iphutha noma inkinga. Uma umsingathi wokudekhoda owabelwana ngawo eyisithikamezo, ukulinganisa okukhishwayo kwenza umthwalo ushelele, kuvimbele ukukhiqizwa okude ekugcwaleni indawo yodwa, futhi kwenza isakazo lifike ngesigqi esilindelekile, esifundekayo esikhundleni sokufika ngamaqoqo. Kuyisinqumo sobunjiniyela esivikelekayo, futhi kwabiza wonke umsebenzisi isikhathi sangempela empendulweni ngayinye.
I-Shannon 3.1 ayinaso isango lesivinini futhi ayinakho ukulinganisa ukubonakaliswa. Amathokheni abhalwa esakazweni lakho njengoba injini iwakhiqiza. Uma injini ikhiqiza ngokushesha, uyibona ikhiqiza ngokushesha. Ayikho ibhafa yokushelelisa phakathi kwemodeli nethemenali yakho, iwindi lakho lengxoxo, noma isifundi sakho se-SSE. Empendulweni ende — ukuhlaziya kwamathokheni angu-2,000, ifayela lekhodi ekhiqiziwe — lokhu kodwa kwakhela ingxenye enkulu yokuthuthuka ozokubona.
Umphumela oqotho: isakazo manje lifika ngamaqoqo. I-speculative decoding (ngezansi) ikhipha amathokheni amukelwayo ngochungechunge olufushane, ngakho umbhalo ungafika ngamaqoqo abonakalayo esikhundleni sokufika ngesigqi segama elilodwa ngesikhathi. Uma wakha i-UI ezungeze isigqi esishelelayo se-3.0, isazosebenza — ukulandelana kwamathokheni nokuqukethwe akuthintekile — kodwa ungase ufune ukubuyisa ukushelelisa kwakho ohlangothini lweklayenti uma uthanda ukubukeka kwesibhali. Sicabanga ukuthi abantu abaningi bangathanda ukubuyiselwa imizuzwana.
Injini uqobo isheshe kakhulu
Ukususa isango kusiza kuphela uma lokho okungemuva kwesango kusheshe. Umthelela wesibili yisitaki senkonzo esichazwe esigabeni 03: izisindo ze-NVFP4 zamabhithi ama-4 kanye ne-speculative decoding kuma-accelerator esizukulwane samanje, angokwemvelo ku-FP4, eqoqweni lethu. Ndawonye lezi zinto ziphakamisa uphahla olwembulwa ukususa isango.
Ukubambezeleka kwempendulo ephelele okuqhathanisiwe, ukuhlolwa kwangaphakathi kwe-Shannon Lab, Mandulo 2026. Ububanzi bubonisa ubude bomyalo nezinga: imiyalo emifushane ku-Lite ihlala eduze nasezingeni eliphansi, ukukhiqizwa okude ku-Pro kuseduze nezinga eliphezulu.
03Yini ngempela eyenziwa yi-speculative decoding
"I-speculative decoding inference" isetshenziswa njengegama lokukhangisa, ngakho nansi indlela esebenza ngayo, ngokucacile.
Imodeli yolimi ngokuvamile ikhiqiza ithokheni elilodwa ekudluleni ngakunye okuya phambili. Lokho kudlula akulawulwa yizibalo kodwa yibandwidth yenkumbulo — izisindo kufanele zifundwe ukuze kubalwe noma yini, futhi ukuzifunda kuthatha isikhathi eside kakhulu kunezibalo. I-GPU ichitha isikhathi sayo esiningi ilinde inkumbulo lapho amayunithi ayo okubala engenzi lutho. Ukukhiqiza amathokheni angu-500 kusho ukukhokha lokho kubambezeleka izikhathi ezingu-500, ngokulandelana.
I-speculative decoding ihlasela ingxenye elandelanayo. Imodeli yohlaka encane, eshibhile ihlongoza uchungechunge olufushane lwamathokheni alandelayo okungenzeka — ake sithi amane noma ayisishiyagalombili. Imodeli eyinhloko bese ihlola lonke lolo chungechunge ekudluleni okukodwa okuhlanganisiwe, okubiza kancane nje kunokuhlola ithokheni elilodwa, ngoba ingxenye ebizayo (ukufunda izisindo) yenzeka kanye noma kunjani. Ithokheni ngalinye elihlongozwayo imodeli eyinhloko evumelana nalo liyamukelwa likhishwe ngokushesha. Ekungavumelaneni kokuqala, uchungechunge luyanqanyulwa bese ukudekhoda okujwayelekile kuqhubeka kusukela lapho.
Ipropathi ebalulekile
I-speculative decoding ilondoloza okukhishwayo. Isinyathelo sokuqinisekisa sakhiwe ngendlela yokuthi uchungechunge olwamukelwayo lusabalale ngendlela efanayo ncamashi nalokho ukusampula kwemodeli eyinhloko okwakuzokwenza. Awutholi impendulo yemodeli yohlaka, futhi awutholi ukulinganisela kwempendulo yemodeli enkulu. Uthola okukhishwa yimodeli eyinhloko, okufinyelelwe ngezinyathelo ezilandelanayo ezimbalwa. Imodeli yohlaka ingathinta isivinini kuphela, ayikwazi ukuthinta okuqukethwe.
Izinga lokwamukelwa yilona elenza umsebenzi. Embhalweni olindelekile — imibhalo ephindaphindwayo, ukwakheka kwekhodi, izixhumanisi zempikiswano — imodeli yohlaka iqagela kahle futhi uchungechunge olude lufika ngesikhathi esisodwa. Kumathokheni anzima ngempela, ukwamukelwa kuyehla futhi uhlelo lubuyela ngokushelelayo ekudekhodeni okujwayelekile kwethokheni elilodwa ngesikhathi. Lokho kungalingani yikho okukufunayo: kusheshisa izingxenye ezilula futhi akuzithinti ezinzima.
Kungani izisindo zamabhithi ama-4 zisendimeni efanayo
Ngoba isithikamezo yibandwidth yenkumbulo, ukwehlisa usayizi wezisindo kuyisikhwepha sesivinini esiqondile, hhayi nje esosayizi wenkumbulo. I-NVFP4 yifomethi yamaphoyinti antantayo yamabhithi ama-4 enokukala okucolekile ngebhloko ngayinye, okuyikho okuyenza igcine ukunemba lapho amasu amadala okwehlisa amanani wamabhithi ama-4 alahlekelwa khona. Cishe ingxenye eyodwa kwezine yamabhayithi okumele afundwe ekudluleni ngakunye kusho isikhathi esincane ngokulinganayo sokulinda inkumbulo, futhi kushiya indawo eningi kakhulu ye-KV cache yomongo omude edingwa yiwindi le-196K.
Lezi zimbili ziyandisana kunokuhlanganiswa nje: amabhayithi ambalwa ekudluleni ngakunye, nokudlula okumbalwa kwithokheni ngalinye elikhishwayo. Yikho okwenza ukusakaza okungavinjiwe, ngesivinini esigcwele kukwazeke ukukhonzwa kunokuba yizindleko obesizodinga ukuzibuyisa ngokulinganisa isivinini — okuyikho ncamashi obekwenziwa ungqimba lokulinganisa ku-3.0.
04Amathokheni angu-196,608: yini evulwa yiwindi eliphindwe ka-6
I-Shannon 3.0 yayinewindi lamathokheni angu-32,768: iseshini yokusebenza, hhayi idokhumenti. Cishe amakhasi angu-70-80 ombhalo, ususa lokho ukuzungeza kokucabanga okukudlayo ngokucabanga kwakho, ususa umyalo wakho wesistimu, ususa ingxoxo kuze kube manje. Umsebenzi wangempela wawushaya lolo donga njalo, futhi izindlela zokusizakala — ukuhlukanisa izingcezu, ukufingqa, ukulanda kokwakho okuqukethwe — zonke zonakalisa into obuzama ukuyilondoloza. Amathokheni angu-196,608 ayisigaba esehlukile senkinga.
Ngokuqondile, lokho cishe kungamakhasi angu-400-500 ombhalo wesiNgisi, noma ikhodibhesi yosayizi omaphakathi kanye nokuhlolwa kwayo ne-README yayo, noma unyaka wamanothi emihlangano yephrojekthi eyodwa, noma iqoqo eliphelele lezivumelwano nazo zonke izethasiselo zalo — kubanjwe engxoxweni eyodwa, kufinyelelwa kuwo ngombuzo owodwa, kungekho ungqimba lokuhlukanisa phakathi kwakho nokuqukethwe.
- Imibuzo ngendawo yokugcina ikhodi yonke. Layisha ikhodi bese ubuza ukuthi kungani inkinga ifinyelela ekukhiqizweni, kunokunamathisela amafayela amathathu obusuvele uwasola. Imodeli ingathola ifayela obungacabanganga ukulifaka.
- Ukuhlaziya idokhumenti elide ngaphandle kokulanda ulwazi. Ukulanda ulwazi kuyisihlungi sokuqala esilahlekisayo esinquma ukuthi imodeli ivunyelwe ukubona ini. Ku-196K uvame ukukweqa bese uvumela imodeli ifunde konke, okususa isigaba sonke sokwehluleka lapho isiqephu esifanele singalandwanga.
- Izingxoxo ezihlala zizwakala. Iseshini yokusebenza ende ayisayilahli ngokuthula isiqalo sayo. Imikhawulo oyibeke emlayezweni wesithathu isasebenza emlayezweni wamashumi ayisishiyagalombili.
- Indawo yokuzungeza kokucabanga. Ukucabanga, ukubhala uhlaka nokubuyekeza kokuzungeza kudla umongo. Ku-3.0 lezo zinyathelo zazincintisana nokuqukethwe kwakho ngesabelo esincane. Ku-3.1 zingena kahle, okuyingxenye yesizathu sokuthi ukuthuthuka kwekhwalithi nokwanda kwewindi kufike ndawonye.
- Okufakwayo okukhulu okuxubile. Izithombe, umbhalo wedokhumenti okhishiwe nekhodi kuthuba olulodwa, ngaphandle kokukhetha ukuthi ikuphi ongakwazi ukukufaka.
Isexwayiso esisodwa esiqotho esisebenza kuyo yonke imodeli yomongo omude, kufaka phakathi eyethu: iwindi elikhulu liyikhono, hhayi isiqinisekiso sokunaka okulinganayo kulo lonke. Ukwakheka kusasiza. Ukubeka umbuzo eduze nesiphetho, ukulebula amadokhumenti akho, nokutshela imodeli ukuthi ifunani konke kuthuthukisa imiphumela ngendlela ekalekayo kumathokheni angu-150K ngendlela okungenzeki ngayo ku-5K.
05I-Lite ne-Pro: shannon-3.1 no-shannon-3.1-pro
Amazinga amabili ahlukile ngokuthi asebenzisa ingxenye engakanani yokuzungeza kokucabanga, futhi lowo yiwona mehluko owodwa obalulekile ekukhetheni phakathi kwawo.
| shannon-3.1 (Lite) | shannon-3.1-pro (Pro) | |
|---|---|---|
| Ukucabanga | Ukudlula okukodwa | Ukuzungeza okuphelele + ukuvuna ulwazi |
| Ukuzibuyekeza | Cha | Yebo |
| Iwindi lomongo | 196,608 | 196,608 |
| Ukusakaza | Isivinini esigcwele, singavinjiwe | Isivinini esigcwele, singavinjiwe |
| Umbono & amadokhumenti | Yebo | Yebo |
| Ithuluzi lokwakha izithombe | Yebo | Yebo |
| Kufanele | Umsebenzi omningi, inani eliphezulu | Imibuzo enzima, uhlaka lokuqala olunganele |
Sebenzisa i-Lite ngokuzenzakalelayo. Ukudlula okukodwa kwemodeli enhle yokucabanga kubhekana neningi lezicelo zangempela, futhi ku-3.1 isheshe ngokwanele ukuthi ukuzungeza akuseyona into oyizwayo ulinde. Yiya ku-Pro lapho umbuzo uwumbuzo lapho impendulo yokuqala ivamise ukuba yiphutha elifundisayo: ukukhetha phakathi kwezakhiwo, ukuhlaziya okuphikisayo, noma yini lapho ungathanda ukuthi ozakwenu onekhono alale phezu kwayo. Isinyathelo se-Pro sokuzibuyekeza asisona umhlobiso — yimodeli ethola amaphutha ayo ngaphambi kokuba kudingeke wena.
06Ukwanda kobuhlakani ngo-32%, nokuthi ukufunda kanjani
Ukuhlolwa kwe-Shannon Lab kubeka i-Shannon 3.1 ekuthuthukeni kobuhlakani ngo-32% ngaphezu kwe-Shannon 3.0. Sifuna ukucacisa ukuthi lokho kuyini nokuthi akuyona ini.
Yinombolo yethu, evela kusethi yethu yangaphakathi yokuhlola, ekalwe emisebenzini esiyibheka njengemelele lokho abantu abakulethayo kulezi zimodeli. Akuyona isilinganiso senkampani yangaphandle, futhi asakhi ithebula lohlu lwabahamba phambili ngenani elilodwa langaphakathi noma sishicilele ukuhlukaniswa ngokwesilinganiso ngasinye — ukuhlukanisa kusho izinga lokuqhathaniswa kwangaphandle isethi yangaphakathi engenalo.
Esizokusho ukuthi ukuthuthuka kuvelaphi, ngoba leyo ngxenye ayifihlekile. Iwindi lomongo elikhulu lisho ukuthi okuqukethwe okuncane kufanele kulahlwe ngaphambi kokuba imodeli icabange ngakho, futhi ukungahlakaniphi okuningi okubonakalayo ezingxoxweni ezinde empeleni ukukhohlwa nje. Isitaki senkonzo esilawulwa yithi sisho ukuthi imodeli isebenza ngokumiswa esasikuhlosile kunokuzenzakalelayo komsingathi. Futhi ukuzungeza kokucabanga — okungashintshile ekuklanyweni — manje kunendawo yokusebenza ngempela ngaphakathi kwewindi kunokucindezelwa uphahla lwe-32K ebeliyabelana nokufakwayo kwakho.
Isilinganiso esiwusizo kakhulu kuwe ngesakho. Thatha umyalo emsebenzini wakho wangempela — hhayi iphazili, owangempela — bese uwusebenzisa ku-shannon-3 bese uwusebenzisa ku-shannon-3.1. Qhathanisa izimpendulo, uzikale ngesikhathi. Izinombolo zethu zichaza isilinganiso kuyo yonke isethi; umyalo wakho yiwona okumele ube ngcono.
07Umbono, amadokhumenti, nokwakha izithombe
I-Shannon 3.1 ifunda izithombe namadokhumenti. Izithombeskrini, imidwebo, izithombe, amakhasi askeniwe, ama-PDF namadokhumenti ombhalo angangena engxoxweni bese kucatshangwa ngawo kanye nakho konke okunye. Ihlanganiswe newindi le-196K, yilokhu okwenza ukusebenza ngedokhumenti eliphelele kusebenzeke: umbiko omude nemidwebo yawo kuthuba olulodwa, ngaphandle kokunquma kusengaphambili ukuthi imodeli ivunyelwe ukubona amaphi amakhasi.
Ukwakha nokuhlela izithombe kuyatholakala engxoxweni njengethuluzi. Cela isithombe bese imodeli ibiza ithuluzi ngaphakathi, engxoxweni efanayo, nomongo wakho konke okuxoxwe ngakho kuze kube manje. Ukuhlela kusebenza ngendlela efanayo — inikeze isithombe, uchaze ushintsho. Ayikho imodi ehlukile okumele ungene kuyo futhi asikho isikhombimsebenzisi esihlukile okumele usifunde.
08Ukubiza i-Shannon 3.1 kusuka ku-API
I-Shannon 3.1 iyatholakala kuzo zontathu izinhlobo ze-API, nokusakaza kuzo zonke. Amamodeli afanayo, izimo ezintathu zesicelo — khetha noma yikuphi okuhambisana ne-SDK osuvele unayo.
| Indawo yokuxhuma | Isimo | Ukusakaza |
|---|---|---|
| /v1/chat/completions | Ihambisana ne-OpenAI | Yebo |
| /v1/messages | Ihambisana ne-Anthropic | Yebo |
| /v1/responses | Responses | Yebo |
{
"model": "shannon-3.1",
"stream": true,
"messages": [
{ "role": "user", "content": "Summarize this contract set and flag anything unusual." }
]
}
Shintsha u-"shannon-3.1" ube ngu-"shannon-3.1-pro" ukuze usebenzise ukuzungeza kokucabanga okuphelele. Uma usuvele ubiza u-shannon-3, ukufuduka kuyi-id yemodeli kuphela: izimo zesicelo nempendulo azishintshile, futhi amaklayenti okusakaza akhona aqhubeka esebenza. Umehluko owodwa wokuziphatha okufanele uwulindele yisakazo elifika ngamaqoqo elichazwe esigabeni 02 — amathokheni afanayo, ngokulandelana okufanayo, afika masinyane futhi ngamaqoqo angalingani kangako.
Uhlu oluphelele lwamapharamitha, ukuqinisekisa, ukusebenza kwamaphutha kanye nendawo yokudlala esebenzisanayo kukwi-dokhumentheshini ye-API. Amanye amakhadi amamodeli nemibhalo yobuchwepheshe akwi-Shannon research.
09Isengavinjelwe, futhi akushintshile kulokho
I-Shannon 3.1 ayinalo ungqimba lokwenqaba futhi ayinakho ukuhlungwa kokuqukethwe kokukhishwayo. Lokhu kufana nakuwo wonke umugqa we-Shannon futhi akushintshanga ngokususwa kuya kuqoqo lethu — uma kukhona, ukulawula isitaki senkonzo kwenza kube lula ukukuqinisekisa, ngoba awukho umsingathi ophakathi onenqubomgomo yawo ohlala phakathi kwemodeli nawe.
Kufanelekile ukukusho ngokucacile ngoba kuyinto esobala okumele umuntu ayibuze ngokwethulwa okushintshe wonke umgudu wokukhishwayo: ukususa ungqimba lokulinganisa isivinini akuchazi ukufaka ungqimba lokulawula esikhundleni sawo. Akukho okuhlolayo, okubhala kabusha, noma okuvimbayo isakazo. Lokho imodeli ekukhiqizayo yikho okufikayo. I-Shannon 3.1, ngokwazi kwethu, yiyona modeli ye-AI engavinjelwe esheshe kunazo zonke etholakalayo enewindi lomongo lobukhulu obungaka — futhi lezi zici ezimbili zihlobene, ngoba zombili zivela ekusebenziseni ingqalasizinda yethu kunokuqasha amandla akhelwe ukuthobela imithetho komunye umuntu.
Ukungavinjelwa akufani nokungaphenduli. Ukusetshenziswa kulawulwa yi-Nqubomgomo yethu Yokusebenzisa Ngokuzibophezela, futhi isibopho esiza nemodeli ezobhekana nokuqukethwe okunzima ukuthi nawe ubhekane nakho ngokuzibophezela.
10Yenza ukuqhathanisa wena ngokwakho
Sonke isimangalo lapha singahlolwa cishe emizuzwini emibili, futhi singathanda ukuthi uhlole kunokuba usikholwe.
- Khetha umyalo emsebenzini owenzayo ngempela. Omude ungcono — usebenzisa kokubili iwindi nokusakaza.
- Wusebenzise ku-
shannon-3. Qaphela ukuthi kuthatha isikhathi esingakanani ukufika kwethokheni lokuqala, nokuthi kuthatha isikhathi esingakanani ukuphela kwempendulo. - Sebenzisa umyalo ofanayo ncamashi ku-
shannon-3.1. Qaphela izinombolo ezimbili ezifanayo. - Bese ufunda zombili izimpendulo, ushaya indiva iwashi, bese unquma ukuthi yiyiphi obungayifuna.
Umehluko wesivinini uzosheshe ubonakale. Umehluko wekhwalithi yiwona okufanele uhlale nawo — ubonakala kakhulu kokufakwayo okude, lapho i-3.0 yayisebenza ngokuthula ngokuqukethwe kwakho okuncane kunalokho obucabanga ukuthi kunjalo.
11Imibuzo ebuzwa njalo
Yini i-Shannon 3.1?
I-Shannon 3.1 iyinguqulo yamanje yomndeni we-Shannon 3. Igcina ukuzungeza kokucabanga okuphindaphindayo okufanayo — cabanga, bhala uhlaka, zibuyekeze, thuthukisa — kodwa ikusebenzisa ngokwemvelo kuqoqo lwethu lwama-GPU esikhundleni somsingathi wangaphandle wokudekhoda. Ukuhlolwa kwe-Shannon Lab kukala ukuthuthuka kobuhlakani ngo-32% nezimpendulo ezisheshayo ngo-10-15x uma kuqhathaniswa ne-Shannon 3.0, iwindi lomongo likhushulwe lisuka kumathokheni angu-32,768 laya kwangu-196,608.
Isheshe kangakanani i-Shannon 3.1 kune-Shannon 3.0?
Isheshe phakathi kuka-10x no-15x ezimpendulweni eziphelele, ngokohlolo lwe-Shannon Lab. Zimbili izinto ezenza lokhu. Okukhishwa yi-Shannon 3.0 kwakudlula kungqimba lokulinganisa isivinini elalunciphisa ngamabomu isakazo; i-Shannon 3.1 ayinaso isango lesivinini futhi ayinakho ukulinganisa ukubonakaliswa, ngakho amathokheni afika kuwe ngesivinini injini ewakhiqiza ngaso. Injini uqobo nayo isheshe kakhulu: izisindo ze-NVFP4 zamabhithi ama-4 kanye ne-speculative decoding kuqoqo lwethu lwama-GPU.
Lingakanani iwindi lomongo le-Shannon 3.1?
Amathokheni angu-196,608, ukukhula okuphindwe ka-6 kunamathokheni angu-32,768 atholakala ku-Shannon 3.0. Lokho cishe kungamakhasi angu-400-500 ombhalo, noma ikhodibhesi yosayizi omaphakathi, kubanjwe engxoxweni eyodwa ngaphandle kokuhlukanisa izingcezu noma ukulanda ulwazi.
Yini i-speculative decoding futhi kungani ibalulekile lapha?
Imodeli encane esheshayo yohlaka ihlongoza uchungechunge lwamathokheni alandelayo okungenzeka, bese imodeli eyinhloko iwaqinisekisa ngokudlula okukodwa okuhlanganisiwe. Amathokheni amukelwayo akhishwa ngokushesha; alahliwe abuyela ekudekhodeni okujwayelekile. Okukhishwayo yilokho imodeli eyinhloko ebizokukhiqiza yodwa, kodwa amathokheni amaningana angafika esinyathelweni ngasinye sokuqinisekisa esikhundleni selilodwa. Yikho okwenza ukusakaza ngesivinini esigcwele kukwazeke ngezindleko kunokuba kube yinkinga yezindleko.
Ingabe i-Shannon 3.1 ayivinjelwe?
Yebo. I-Shannon 3.1 ayinalo ungqimba lokwenqaba futhi ayinakho ukuhlungwa kokuqukethwe okusetshenziswa kokukhishwayo kwayo, njengoba kunjalo nakuwo wonke umugqa we-Shannon. Ukususa ungqimba lokulinganisa isivinini akuzange kufake ungqimba lokulawula esikhundleni sawo — isakazo olitholayo yikho okukhishwa yimodeli.
Yiziphi ama-id emodeli futhi ngingayisebenzisa kuphi i-Shannon 3.1?
I-shannon-3.1 yizinga le-Lite kanti i-shannon-3.1-pro yizinga le-Pro. Zombili ziyatholakala engxoxweni nakuzo zontathu izinhlobo ze-API: /v1/chat/completions (yakhelwe i-OpenAI), /v1/messages (yakhelwe i-Anthropic), kanye ne-/v1/responses. Ukusakaza kuyasebenza kuzo zontathu.
Uyini umehluko phakathi kwe-shannon-3.1 ne-shannon-3.1-pro?
I-Lite isebenzisa ukudlula okukodwa kokuzungeza kokucabanga: iyacabanga, bese iphendula. I-Pro isebenzisa ukuzungeza okuphelele — cabanga, bhala uhlaka, zibuyekeze, thuthukisa — nesinyathelo sokuvuna ulwazi ngaphambi kokubhala uhlaka. I-Lite iyisisekelo esifanele emsebenzini omningi; i-Pro ingeyemibuzo lapho impendulo yokuqala ivamise ukungabi engcono khona.
Zama i-Shannon 3.1
Ukuzungeza kokucabanga okufanayo. Iwindi eliphindwe kasithupha. Alikho isango lesivinini.
Qala Ukuxoxa Funda Imibhalo ye-APIshannon-3.1 · shannon-3.1-pro · ukusakaza kuzo zontathu izinhlobo
Izibalo zokusebenza nezobuhlakani kulesi sihloko yizilinganiso ze-Shannon Lab uqobo ezivela ekuhloleni kwangaphakathi, Mandulo 2026, futhi zishiwo njengezibalo zomkhiqizo kunokuba yimiphumela yezilinganiso zenkampani yangaphandle. Iwindi lomongo, ama-id emodeli, nokutholakala kwe-API yizincazelo zomkhiqizo. I-Shannon AI iqhutshwa yi-Shannon Lab LLC, New Mexico, USA. Okuhlobene ongakufunda: Shannon 3 · Inkomba yocwaningo lwe-Shannon · Idokhumentheshini ye-API.