We do not hold or access any user's data, and we do not suspend accounts unless a lawful authority requires an enforcement action.
I-NVFP4 e-Quantized - I-AI Yebhizinisi Engabizi Kakhulu

ShannonLite 1.6

I-AI yebhizinisi engabizi kakhulu enikwa amandla yi-Mistral Large 3nge-amapharamitha angu-675B esewonkekanyeamapharamitha asebenzayo angu-41Bngesakhiwo se-Mixture-of-Experts esinezingxenye ezincane. Iqeqeshwe ngemuva ku-imiphumela engu-2,500 ye-Claude Opus 4.5ukuze kulandelwe imiyalelo ngendlela engavamile.I-NVFP4 quantizationyenza kube nokwenzeka ukusetshenziswa kwe-node eyodwa ku-H100s noma A100s.

675B
Amapharamitha Ewonke
41B
Amapharamitha Asebenzayo
NVFP4
I-Quantization
256K
Ulwazi Olungemuva
2.5B
I-Vision Encoder
Uhlelo lwe-Lite
Shannon Lite 1.6
v1.6.0-lite-nvfp4
Imininingwane Yobuchwepheshe:
Imodeli Eyisisekelo Mistral Large 3
Isakhiwo I-Granular MoE
Amapharamitha Ewonke 675B
Amapharamitha Asebenzayo 41B
I-Quantization NVFP4
Ngemuva Kokuqeqeshwa Claude Opus 4.5
Amasampula Okuqeqesha 2,500

Mistral Large 3: I-Granular Mixture-of-Experts

I-Shannon Lite 1.6 yakhelwe ku-Mistral Large 3, imodeli ye-Mixture-of-Experts enezingxenye ezincane, eyingqayizivele, enezindlela eziningi eyakhelwe kusukela phansi ukuze ithembeke, iqonde ulwazi olungemuva olude, futhi isebenze kahle ezingeni lokukhiqiza. Inguqulo eqeqeshwe ngemuva, efundisayo, ilungiswe kahle ukuze isetshenziswe ezingxoxweni, ezinhlelweni ze-agentic, nasezimweni zokusebenzisa ezisekelwe emiyalweni.

673B

Imodeli Yolimi

Isakhiwo se-Granular MoE ngamapharamitha asebenzayo angu-39B ngokuphasa ngakunye phambili

2.5B

I-Vision Encoder

I-encoder enezindlela eziningi ehlanganisiwe yokuhlaziya izithombe nokuqonda okubonakalayo

256K

Iwindi Lolwazi Olungemuva

Ulwazi olungemuva olwandisiwe lokuqonda amadokhumenti ngokuphelele kanye ne-RAG

12+

Izilimi

IsiNgisi, IsiFulentshi, IsiSpanishi, IsiJalimane, IsiShayina, IsiJapane, IsiKorea, Isi-Arabhu, nokunye okuningi

Ukusetshenziswa Kwebhizinisi Okungabizi Kakhulu

I-Shannon Lite 1.6 isebenzisa ubuchwepheshe be-NVFP4 (4-bit floating point) quantization be-NVIDIA ukuze yehlise kakhulu izidingo zememori ngenkathi igcina ikhwalithi yemodeli. Sebenzisa i-AI esezingeni eliphezulu engqalasizinda ye-GPU efinyelelekayo ngaphandle kobunzima be-multi-node.

💰

Izindleko Zengqalasizinda Ezincishisiwe

I-NVFP4 quantization yehlisa ukusetshenziswa kwememori cishe ngokuphindwe ka-4 uma kuqhathaniswa ne-BF16, okwenza kube nokwenzeka ukusetshenziswa kuma-GPU ambalwa futhi kunciphise kakhulu i-TCO ye-AI yebhizinisi.

Ukusetshenziswa Kwe-Node Eyodwa

Sebenzisa imodeli ephelele yamapharamitha angu-675B ku-node eyodwa yama-H100s noma ama-A100s. Akukho ukuhlela okuyinkimbinkimbi kwe-multi-node, ukuncipha kwezindleko zenethiwekhi, ukusebenza okulula.

Ikhwalithi Yemodeli Egciniwe

Amasu athuthukile e-quantization agcina ukusebenza kwemodeli ekucabangeni, ekulandeleni imiyalelo, nasemisebenzini enezindlela eziningi ngokuncipha okuncane kwekhwalithi.

I-Claude Opus 4.5 Knowledge Distillation

I-Shannon Lite 1.6 iqeqeshwe ngemuva ngokucophelela kusetshenziswa imiphumela engu-2,500 ekhethwe ngokucophelela evela ku-Claude Opus 4.5, imodeli ye-Anthropic ekwazi kakhulu. Le ndlela ye-knowledge distillation ithwebula amaphethini okucabanga athuthukile, ukuhunyushwa kwemiyalelo okucashile, nekhwalithi ephezulu yempendulo.

Mistral Large 3 Instruct 2512 Foundation

Yakhelwe ku-Mistral's state-of-the-art Instruct model (inguqulo 2512) ngokunemba kwe-BF16. Lesi sisekelo sinikeza amakhono asezingeni eliphezulu aklanyelwe abasizi bezinga lokukhiqiza, izinhlelo ezithuthukisiwe zokubuyisa ulwazi, imisebenzi yesayensi, kanye nezinhlelo zokusebenza zebhizinisi eziyinkimbinkimbi.

Isisekelo se-BF16 I-Instruct Elungisiwe Ilungele Ukukhiqiza Ilayisense ye-Apache 2.0

Claude Opus 4.5 Output Distillation

Iqeqeshwe ngemuva ngemiphumela engu-2,500 esezingeni eliphezulu evela ku-Claude Opus 4.5, ithwebula amakhono okucabanga athuthuke kakhulu e-Anthropic. Idatha ekhethiwe igxile ekulandeleni imiyalelo eyinkimbinkimbi, ekuqondeni okucashile, nasekukhiqizeni izimpendulo ezisezingeni eliphezulu kuzo zonke izizinda ezahlukene.

Amasampula angu-2,500 Idatha Ekhethiwe Ukugxila Kwekhwalithi Izizinda Ezihlukahlukene

Inqubo Yokulinganisa ye-NVFP4

Ukulinganisa okuthuthukisiwe kwe-NVIDIA FP4 okusetshenziswe ngemuva kokuqeqeshwa ukuze kuncishiswe ukusetshenziswa kwememori ngenkathi kugcinwa ikhwalithi yemodeli. Kulinganiswe ngokukhethekile izisindo eziqeqeshwe ngemuva ukuze kugcinwe ukudluliswa kolwazi lwe-Claude Opus 4.5 namakhono okulandela imiyalelo.

NVFP4 Ukunemba Okungama-4-bit Kulinganisiwe Ikhwalithi Igciniwe

Ukuhlola Nokugunyaza

Ukuhlola okuphelele kuzo zonke izilinganiso zokulandela imiyalelo, imisebenzi yokucabanga, nezimo zangempela zebhizinisi. Kugunyazwe ukuziphatha okungaguquki kuzo zonke izizinda, imiphumela ezinzile, nokusebenza okuthembekile ezindaweni zokukhiqiza.

Kulinganiswe Ngokwezinga Kuzo Zonke Izizinda Kugunyazwe Ukukhiqiza Imiphumela Ezinzile

Izinketho Eziguquguqukayo Zokuthunyelwa Kwe-GPU

I-Shannon Lite 1.6 enokulinganisa kwe-NVFP4 yenza ukuthunyelwa okungabizi kakhulu kumalungiselelo e-NVIDIA GPU ajwayelekile embonini, okwenza i-AI esezingeni eliphezulu ifinyeleleke ekuthunyelweni kwamabhizinisi ngaphandle kokudinga amaqoqo amaningi abizayo.

NVIDIA H100 SXM

Ukusebenza okuhle kakhulu ngezakhiwo ze-Hopper nememori ye-HBM3

I-Node Eyodwa (8x H100)
Ukunemba kwe-NVFP4
80GB HBM3 nge-GPU ngayinye
Ukudlula Okukhulu Kakhulu

NVIDIA A100 SXM

Ukuthembeka okufakazelwe kuma-GPU ezakhiwo ze-Ampere

I-Node Eyodwa (8x A100)
Ukunemba kwe-NVFP4
80GB HBM2e nge-GPU ngayinye
Okungabizi Kakhulu

Shannon Cloud

Ukuthunyelwa okuphethwe ngokugcwele ngaphandle kwengqalasizinda

Ukufinyelela Ngokushesha
Ukuzenzakalela Kokulinganisa
Ilungele i-REST API
99.9% SLA

Izici ze-AI Ezilungele Ibhizinisi

I-Shannon Lite 1.6 iletha amakhono asemngceleni azuzwe ku-Mistral Large 3 futhi athuthukiswa nge-Claude Opus 4.5 ngemuva kokuqeqeshwa, elungiselelwe imisebenzi yokukhiqiza kuzo zonke izimo zebhizinisi ezahlukahlukene.

Umbono Wezindlela Eziningi

I-encoder yombono eneparamitha engu-2.5B ehlanganisiwe yenza ukuhlaziya isithombe, ukuphendula imibuzo ebonakalayo, nokuqonda amadokhumenti ngezithombe.

Ubuhle Bezilimi Eziningi

Ukusekelwa kwendabuko kwezilimi ezingaphezu kweziyi-12 kuhlanganisa isiNgisi, isiFulentshi, iSipanishi, isiJalimane, isiNtaliyane, isiPutukezi, isiDashi, isiShayina, isiJapane, isiKorea, nesi-Arabhu.

🤖

Amakhono Okuba Umenzeli

Izici zomenzeli ezihamba phambili ezinezingcingo zomsebenzi wendabuko kanye nokuphuma kwe-JSON okuhlelekile ukuze kusetshenziswe amathuluzi azimele kanye nokuzenzakalela kokugeleza komsebenzi.

Ukulandela Okusheshayo Kwesistimu

Ukulandela okuqinile nokusekelwa kwemiyalo yesistimu, okwenza ukulawula kokuziphatha okunembile nokugcinwa komuntu okungaguquki.

Umongo Omude Ongu-256K

Iwindi lomongo elinwetshiwe lokuqonda amadokhumenti okuphelele, izingxoxo ezinwetshiwe, nokukhiqizwa okuthuthukisiwe ngokubuyiswa (RAG).

🔧

Ukubiza Umsebenzi Wendabuko

Ukusekelwa kokubiza umsebenzi okwakhelwe ngaphakathi ngokuphuma kwe-JSON okuthembekile ukuze kuhlanganiswe kalula namathuluzi angaphandle, ama-API, nezinsizakalo.

Ilungiselelwe Imisebenzi Yokukhiqiza

Ngokusebenza okunamandla komongo omude, ukuziphatha okuzinzile nokungaguquki kuzo zonke izizinda, i-Shannon Lite 1.6 ihamba phambili kuzo zonke izimo zebhizinisi nezocwaningo ezahlukahlukene.

📄

Ukuqonda Amadokhumenti Amade

Hlaziya futhi uhlaziye amadokhumenti amakhulu, izinkontileka, imibiko, namaphepha ocwaningo ngewindi lomongo elingu-256K

🤖

Abasizi be-AI Bokukhiqiza

Nika amandla abasizi be-AI abasebenzisa nsuku zonke ngezimpendulo ezithembekile, ezingaguquki, nokulandela imiyalelo okuqinile

🔧

Ukuhamba Komsebenzi Womzeli

Ukusetshenziswa kwamathuluzi asezingeni eliphezulu nokubiza umsebenzi ukuze kwenziwe umsebenzi ozimele kanye nokuzenzakalela kokugeleza komsebenzi

🏢

Umsebenzi Wolwazi Lwebhizinisi

Ukuhamba komsebenzi okuyinkimbinkimbi kwebhizinisi okudinga amakhono e-AI asemngceleni anokuphuma okungaguquki, okuthembekile

💻

Umsizi Ojwayelekile Wokubhala Ikhodi

Ukukhiqizwa kwekhodi, ukulungisa amaphutha, ukubhala amadokhumenti, nosizo lokuthuthukiswa kwesoftware kuzo zonke izilimi eziningi

Ucwaningo Lwesayensi

Usizo locwaningo, ukubuyekezwa kwezincwadi, ukucubungula umsebenzi wesayensi, nokukhiqizwa kwemibono

Ukukhiqizwa Okuthuthukisiwe Ngokubuyiswa

Ukusebenza okuhle kakhulu kwezinhlelo ze-RAG ngokuhlanganiswa komongo okuthembekile nokuhlanganiswa kokubuyiswa okunembile

🌍

Izinhlelo Zokusebenza Zezilimi Eziningi

Izinhlelo zokusebenza zebhizinisi zomhlaba wonke ezidinga ikhwalithi engaguquki kuzo zonke izilimi ezingaphezu kweziyi-12 ezisekelwayo

I-Shannon Lite vs I-Shannon Pro

Khetha imodeli efanele ye-Shannon ngezidingo zakho. I-Shannon Lite inikeza ukuthunyelwa kwebhizinisi okungabizi kakhulu, kanti i-Shannon Pro inikeza amakhono amakhulu ngokucabanga okuthuthukisiwe kwe-chain-of-thought nokusekelwa kwamakhono.

Isici Shannon Lite 1.6 Shannon Pro 1.6
Imodeli Eyisisekelo Mistral Large 3 (675B) Mistral Large 3 (675B)
Amapharamitha Asebenzayo 41B (I-MoE Enemininingwane) 41B (I-MoE Enemininingwane)
Ukunemba NVFP4 (4-bit) Okugcwele i-BF16 (16-bit)
Idatha Yangemuva Kokuqeqeshwa Imiphumela engu-2,500 ye-Claude Opus 4.5 Imikhondo Yokucabanga ye-KIMI K2
Indlela Yangemuva Kokuqeqeshwa Ukulungisa Okuncane Okugadiwe I-GRPO (Ukuthuthukiswa Kwenqubomgomo Ehlobene Neqembu)
Imodi Yokucabanga Okujwayelekile Imikhondo Yokucabanga Elandelanayo
Ukusekelwa Kwamakhono - OkwabaPro KuphelaAmakhono Omdabu
Ukuthunyelwa H100/A100 (Inodi Eyodwa) B200/H200 (FP8)
Okungcono Kakhulu Ku- I-AI Yebhizinisi Engabizi Kakhulu Amandla Aphezulu + Ukucabanga

Udinga Ukucabanga Namakhono Athuthukile?

I-Shannon Pro 1.6 inezici ze-KIMI K2 Thinking Traces ngokuqeqeshwa kwe-GRPO ukuze kube nokucabanga okusobala okuyimikhondo elandelanayo, kanye nokusekelwa kwamakhono omdabu emisebenzini ye-AI eyenziwe ngokwezifiso.

Hlola i-Shannon Pro

Izwa i-Shannon Lite 1.6

Amakhono e-AI asezingeni eliphezulu anokulinganisa kwe-NVFP4 okungabizi kakhulu. Thumela engqalasizindeni ye-H100 noma i-A100 ukuze uthole ukusebenza kwezinga lebhizinisi ngezindleko ezifinyelelekayo.

All research links