I-API ye-DeepSeek Flash Engavinjelwe

Amaqembu amaningi ahlola i-API ye-LLM alungiselela into engalungile. Ahlola imiyalelo engamashumi amabili enzima kunayo yonke angayicabanga, akhethe onqobile, bese ethola emasontweni ayisithupha kamuva ukuthi amaphesenti angamashumi ayisishiyagalolunye nanhlanu ethrafikhi yawo ngumsebenzi omfushane, ohlelekile, ophindaphindekayo imodeli ebiza ingxenye eyodwa kwezine ebingawuphatha ngendlela efanayo. I-DeepSeek-V4-Flash ikhona ngenxa yalawo maphesenti angamashumi ayisishiyagalolunye nanhlanu. Leli khasi limayelana nokwazi ukuthi nini ukukuwo — nokuthi nini ungekho.

01Yini oyitholayo ngempela

Isakhi	DeepSeek-V4-Flash-0731 · W4A16-AUTOROUND-REAP
I-id yemodeli	DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP
Umongo / okukhishwayo okuphezulu	amathokheni angu-262,144 / amathokheni angu-16,384
Intengo nge-1M (engenayo / ephumayo)	$0.50 / $2.00
Ukusakaza · ukubiza amathuluzi · ukucabanga	Yebo — zonke izilimi ezintathu
Okukhishwayo okuhlelekile	json_schema kanye no-response_format
Okufakwayo okuyisithombe (umbono)	Cha — umbhalo kuphela
Ungqimba lokwenqaba / isihlungi sokukhishwayo	Alukho
Ikhonzwa ku-	iqoqo lama-GPU le-Shannon

Izilimi ezintathu ze-API zifinyelela ezisindweni ezifanayo: /v1/chat/completions (esimweni se-OpenAI), /v1/messages (esimweni se-Anthropic), kanye no-/v1/responses. Ukusakaza kuyasebenza kuzo zontathu. Ireferensi ephelele yamapharamitha iku-mibhalo ye-API.

02Ivelaphi: uzalo lwe-DeepSeek

I-DeepSeek-AI ichithe izizukulwane ezinhlanu zamamodeli phezu kwento eyodwa: ukukhipha ukuziphatha okusemkhawulweni emodelini eshibhe ukukhonzwa, hhayi nje eshibhe ukuqeqeshwa. Cishe sonke isinqumo sokwakhiwa kulolu hlu sibuyela kulokho, futhi ezibalulekile zishicilelwe. Ezintathu zichaza lokho okubizayo uma ushaya lesi siphetho.

DeepSeekMoE — ochwepheshe abancane kanye noyedwa owabelwanayo

Ingqimba ye-Mixture-of-Experts ithatha indawo yenethiwekhi eyodwa enkulu ye-feed-forward ngezinye eziningi ezincane kanye ne-router ekhetha abambalwa kuthokheni ngalinye. Inguqulo ye-DeepSeek ihlukanisa ochwepheshe ngokuncane kunokuhleleka okujwayelekile kochwepheshe abayisishiyagalombili, ngakho ama-FLOP asebenzayo afanayo athenga inhlanganisela eqondile kakhulu, futhi ihlukanisa uchwepheshe owabelwanayo ithokheni ngalinye elidlula kuye, ngakho ulwazi olujwayelekile alufundwa kabusha ngokuphindaphindiwe ngaphakathi kochwepheshe ngamunye. I-DeepSeek-V3 yathumela lokho njengochwepheshe abangu-256 abaqondiswayo kanye nochwepheshe oyedwa owabelwanayo, kusetshenziswa abangu-8 abaqondiswayo kuthokheni ngalinye — amapharamitha angu-671B esewonke, angu-37B asebenzayo — kanye nokulinganiswa komthwalo okungenalahleko elekelelayo: kunenani lesijeziso elehlisa ikhwalithi, kunokuchezuka kochwepheshe ngamunye okusunduza amaphuzu okuqondisa phezulu noma phansi esinyathelweni ngasinye njengoba lowo chwepheshe esebenza engaphansi noma ngaphezu komthwalo.

MLA — ukunaka okulingana enkumbulweni

Enye ingxenye yenkinga yezindleko zokukhonza yi-KV cache, ekhula ngokulingana nomongo futhi ebusa inkumbulo ebudeni obude bochungechunge. I-Multi-head Latent Attention, eyethulwa ku-DeepSeek-V2, icindezela izikhiye namanani ndawonye kube yi-latent vector ephansi bese ikhesha leyo, yakha kabusha u-K/V osebenzisekayo ngesikhathi sokusebenza. Amanani e-V2 ashicilelwe we-DeepSeek, uma eqhathaniswa ne-DeepSeek 67B yabo eminyene yangaphambili, abika ukwehla kwe-KV cache okungu-93.3% kanye nokukhishwa okuphezulu kokukhiqiza okuphindwe ka-5.76. Yingakho iwindi lamathokheni angu-262K liyinto ejwayelekile kulo mndeni kunokuba yisengezo esibizayo.

Ukunaka okusabalele — nesibonelo samathokheni ashibhile

I-V3.2-Exp yengeza i-DeepSeek Sparse Attention: inkomba elula inikeza amathokheni omongo amaphuzu ngokumelene nombuzo, bese isikhethi sigcina kuphela lawo anamaphuzu aphezulu. Ingxenye ethakazelisayo yisiqinisekiso, hhayi indlela. I-DeepSeek yashicilela ukuqhathaniswa nge-V3.1-Terminus lapho i-MMLU-Pro ilingana ku-85.0, i-AIME 2025 yenyuka (88.4 → 89.3), i-SWE-bench Verified yehla kancane (68.4 → 67.8) — ikhwalithi ilinganile — futhi banciphisa amanani abo e-API ngaphezu kuka-50% ngalo lolo suku. Ukwakhiwa, okubikwe ngeqiniso, kuguqulwa ngqo kwaba yintengo. I-Flash yileyo mpikiswano ithathwa isinyathelo esisodwa ngaphezulu.

Isizukulwane se-V4

I-V4, eyabonwa kuqala ngo-Ephreli 2026 ngaphansi kwelayisensi ye-MIT, ithatha indawo yesitaki sokunaka futhi ngenhlanganisela ye-Compressed Sparse Attention ne-Heavily Compressed Attention, iqeqeshwa nge-Muon optimizer emathokhenini angaphezu kwangu-32T, futhi igcina inhloso ye-V3 ye-multi-token-prediction — ephinda isebenze njenge-speculative decoding, futhi i-DeepSeek eyayikala ku-izinga lokwamukelwa kwethokheni lesibili elingu-85–90% ukuze kusheshiswe ukudikhoda okuphindwe cishe ka-1.8 ku-V3. Isifinyezo se-V4 sibika i-V4-Pro isebenza ngo-27% wama-FLOP e-V3.2 okuqagela kwethokheni elilodwa nango-10% we-KV cache yayo. Umndeni uveza amazinga amathathu omzamo wokucabanga, okuyikho ipharamitha yethu ye-reasoning_effort ekhomba kukho.

Imithombo: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · DeepSeek-V3 Technical Report (arXiv 2412.19437) · ikhadi lemodeli le-DeepSeek-V3.2-Exp · DeepSeek-V4 (arXiv 2606.19348). Wonke amanani abikwe yi-DeepSeek.

03Kusho ukuthini ngempela u-"Flash"

Yileli phuzu iningi lokuqhathanisa elilifaka eceleni, ngakho kufanele lishiwo ngokuqondile: i-Flash ye-DeepSeek akuyona i-Pro ekaliwe, ethenwe noma ecutshunguliwe. Iyimodeli ye-Mixture-of-Experts encane eqeqeshwe ngokwehlukene enokuqeqeshwa kwayo kwangaphambili. Iphepha lokubonwa kwe-V4 linikeza osayizi ababili:

Amanani asuka ephepheni lokubonwa kwe-DeepSeek-V4. Amakheshiphoyinti asemthethweni esiwakhonzayo — u-V4-Flash-0731 no-V4-Pro-0813 — abala izinombolo eziphelele ezinkulu kancane ezingu-304B nangu-1.7T emakhadini awo emodeli; i-DeepSeek ayiwaphindi isho amapharamitha asebenzayo alokho kwakhiwa.

Izinombolo ezimbili zenza umsebenzi lapha. Amapharamitha esewonke abeka isikhala senkumbulo, esinquma ukuthi ikhophi yemodeli ithatha ama-hardware amangakanani. Amapharamitha asebenzayo abeka izibalo kuthokheni ngalinye, ezinquma ukukhishwa. I-Flash cishe incane ngokuphindwe ka-5.6 kokokuqala nangokuphindwe ka-3.8 kokwesibili — futhi kunento encane efanele ukuqashelwa kulowo mehluko: i-Flash isebenzisa angu-13 kwayizigidigidi ezingu-284 zamapharamitha, cishe u-4.6%, kanti i-Pro isebenzisa angu-49 kwayizigidigidi ezingu-1,600, cishe u-3.1%. I-Flash yiyona modeli engasabalele kangako. Akuyona nje i-Pro enochwepheshe abambalwa abavuliwe; iwumklamo ominyene kakhulu, olinganiswe ngokwehlukile, ochitha ingxenye enkulu yawo kulo lonke ithokheni elikhiqizayo. Lokho kuyisinqumo sobunjiniyela sangempela, hhayi ukwehliswa.

Ashoni amanani e-DeepSeek uqobo

Womabili amakhadi emodeli asemthethweni ashicilela amaphuzu ekuhlolweni okufanayo kwama-ejenti nokobunjiniyela besofthiwe, okwenza lokhu kube ukuqhathanisa okuhlanzeke ngokungajwayelekile — ilebhu efanayo, umshini wokuhlola ofanayo, osayizi ababili:

Ukuhlolwa (okubikwe yi-DeepSeek)	V4-Flash-0731	V4-Pro-0813
Terminal Bench 2.1	82.7	87.9
Toolathlon-Verified	70.3	74.1
DeepSWE	54.4	62.7
NL2Repo	54.2	61.5
DSBench-FullStack	68.7	71.1

I-Pro inqoba kuwo wonke umugqa. Lokho kuyisihloko esiqotho, futhi singathanda ukuthi usifunde lapha kunokusithola ekukhiqizeni. Kodwa funda imikhawulo, ngoba yiyo oyithengayo ngempela. Ku-Terminal Bench 2.1 umehluko ungamaphuzu angu-5.2 emodelini enengxenye eyodwa kwezingamashumi amabili nanhlanu yamapharamitha; ekusetshenzisweni kwamathuluzi ungamaphuzu angu-3.8. Umehluko uyakhula lapho umsebenzi ujula khona — amaphuzu angu-8.3 ku-DeepSWE, angu-7.3 ku-NL2Repo, womabili avuza ukucabanga kwekhodi okulandelanayo okunezinyathelo eziningi. Yileso isimo sokushintshana ethebhulini elilodwa: i-Flash ivala ingxenye enkulu yebanga emisebenzini enemikhawulo bese idedela indawo kuleyo evulekile.

Ikhadi le-Flash le-DeepSeek uqobo lenza iphuzu ngokuqonde ngokwengeziwe: i-Flash-0731 idlula ukubonwa kwe-V4-Pro ekuhlolweni okubaliwe naphezu kwenani layo elincane kakhulu lamapharamitha asebenzayo. Imodeli encane entsha enqoba enkulu endala kuyisimo esijwayelekile kulo mkhakha, futhi yisona sizathu sokuthi u-"izinga elikhulu" akufanele abe yimpendulo yakho ezenzakalelayo.

04Ukufunda isijobelelo: W4A16-AUTOROUND-REAP

Ama-id ethu emodeli asho ngokuzwakalayo uhlelo lwawo lokucindezela, ngoba yiyo into ethinta kakhulu ikhwalithi yokukhishwayo futhi cishe akekho oyidalulayo. Lesi sihlukaniseka sibe yizimangalo ezintathu.

W4A16 — izisindo zamabhithi amane, ukusebenza kwamabhithi ayi-16

Izisindo zigcinwa emabhithini amane ngasinye; ukusebenza kugeleza ngokunemba kwamabhithi ayi-16. Lokhu kungukukalwa kwezisindo kuphela, futhi kuqondise kumkhawulo wangempela ekukhiqizeni amathokheni, okuyibhandiwidi lenkumbulo kunezibalo. Ukukala nokusebenza bekungathenga isivinini esincane esengeziwe kodwa kubeke engozini iwa elikhulu kakhulu lokunemba; ukukushiya emabhithini ayi-16 kuyisinqumo esinespoki, esethembeke kakhulu. Lokhu kubhalwa kuyifushaniso lezinhlelo zokukhonza — imibhalo yokukala ibhala into efanayo njengo-W4 nosayizi weqembu.

AUTOROUND — ukugoba okufundiwe, hhayi ukugoba okuseduze

Ukukalwa okulula kugoba sonke isisindo siye kwinani eliseduze elingamelwa. Lokho ngokusobala akulona okungcono kakhulu: ukugoba isisindo esisodwa ngendlela "engalungile" kungakhansela iphutha elethulwe ngomakhelwane baso, futhi ngezisindo eziyizigidigidi lezo zinto ezenzekayo ziyabusa. I-AutoRound, evela ku-Intel, iphatha isinqumo sokugoba uqobo njengento okufanele ifundwe — ukushukumiswa okuqeqeshekayo esisindweni ngasinye okuvinjelwe ku-±0.5, ngokwanele ncamashi ukuguqula noma yisiphi isinqumo esisodwa sokugoba futhi hhayi ngaphezulu, kanye namabanga okunquma afundekayo, alungiselelwa nge-signed gradient descent ngokumelene nenhloso yokwakha kabusha okukhishwayo ngamabhlogo esethini encane yokukalibhula. Izinyathelo ezingamakhulu amabili, bese kuphela.

Ukuqhathaniswa okushicilelwe kwe-Intel kukhombisa i-AutoRound inqoba i-GPTQ ezimweni ezingu-30 kwezingu-32 ezihloliwe, i-AWQ kwezingu-27 kwezingu-32 kanye ne-OmniQuant kwezingu-29 kwezingu-32 kwizilungiselelo zezisindo kuphela zamabhithi amane, ngenzuzo ngaphezu kokugoba okulula kwe-round-to-nearest engamaphuzu okunemba angu-6.9 kuya kwangu-33.2 emabhithini amabili. Akukho lutho kwakho okubiza lutho ngesikhathi sokuqagela: kungukukhethwa okungcono kwezinombolo, okwenziwe kanye, ngaphandle kwe-inthanethi.

REAP — ukususa ochwepheshe okungekho oqondisa kubo

I-REAP (Router-weighted Expert Activation Pruning), evela ku-Cerebras Systems nase-University of Calgary, ingukucindezela kokushaya kanye kwamamodeli e-MoE. Inikeza uchwepheshe ngamunye amaphuzu ngokuthi i-router iqondisa kuye ngamandla amangakanani nangokuvamile kangakanani, kuphindwe ngobukhulu bokukhishwayo kwalowo chwepheshe, bese isusa ngokuphelele ochwepheshe abangabalulekile kunabo bonke — kungekho ukuqeqeshwa kabusha ngemuva.

Okutholakele okuphakathi kwephepha yikho okuwusizo: emazingeni aphezulu okucindezela, ukuthena ochwepheshe kudlula ukubahlanganisa. Ukuhlanganisa ochwepheshe abahlukene ngokusebenza kubanga lokho ababhali abakubiza ngokuthi ukuwa kwesikhala sokusebenza — i-router ilahlekelwa ukulawula okuncane, okuxhomeke kokufakwayo okwenza i-MoE ifanelwe ukwakhiwa, bese isala iqondisa phakathi kwezilinganiso ezingashintshi. Amanani abo emamodelini asuka ku-20B kuya ku-1T: ekucindezelweni kochwepheshe okungu-25% emisebenzini yekhodi, i-REAP ilahlekelwa u-2.8% wesilinganiso sokunemba lapho izisekelo zokuhlanganisa zilahlekelwa ngaphezu kuka-5%; ku-50%, i-REAP ilahlekelwa u-8.0% lapho ukuhlanganisa kulahlekelwa ngaphezu kuka-20%. Ku-Qwen3-Coder-480B ekuthenweni kwama-50% ukwehla kungu-1.4%, ngokumelene no-39.3% wokuthena okulula okusekelwe kumvamo. I-Cerebras ishicilele amakheshiphoyinti e-REAP ohlwini lwe-DeepSeek phakathi kwamanye.

Kungani amabhithi amane lapha namabhithi amathathu ku-Pro

Womabili amazinga kufanele alingane futhi ahlale esheshe ohlotsheni olufanayo lwama-hardware. I-Pro iphethe amapharamitha ayizigidigidi eziyinkulungwane namakhulu ayisikhombisa, ngakho idinga ukwelashwa okuqinile kwamabhithi amathathu ukuze ilingane nhlobo. I-Flash iphethe cishe izigidigidi ezingamakhulu amathathu, ngakho ayikudingi — futhi ibhithi elisindisayo lichithwa ekwethembekeni esikhundleni sokulingana.

Lelo bhithi elengeziwe libaluleke ngaphezu kwalokho izibalo zalo ezikushoyo, ngoba umonakalo wokukalwa awuqondile. Ekukalweni kwe-WikiText-2 kwephepha le-AWQ, i-Llama-2-7B isuka ku-5.47 perplexity ku-FP16 iye ku-5.60 emabhithini amane — isijeziso esingu-2.4% — kodwa iye ku-6.24 emabhithini amathathu, isijeziso esingu-14%. Lokho cishe kuphindwe kayisithupha ukwehla ngokucindezelwa okwengeziwe okungu-25%, futhi iphethini efanayo iyabambeka ku-70B. Iphepha le-GPTQ liqonde ngaphezulu: emabhithini amathathu, ukugoba okulula kwe-round-to-nearest kuyawa emamodelini ohlobo lwe-175B, kuye ezinkulungwaneni ze-perplexity, kuyilapho kusasebenziseka kahle emabhithini amane. Ngaphansi kwamabhithi amane, i-algorithm enhle yokukala iyeka ukuba yinto enhle nje ibe yiyona nto kuphela ebamba imodeli ndawonye.

Ngakho ukufundwa okuqotho kwalezi zijobelelo ezimbili yilokhu: i-Flash ihlala eduze nemodeli yayo engacindezelwe kunokuba i-Pro ihlale eduze neyayo. I-Pro isengumodeli enamandla kakhulu — ithebula lokuhlolwa elingenhla likunquma lokho. Ivele yadedela okuningi ukuze ifike lapha, futhi ukhokhela lelo lungelo.

Imithombo: AutoRound — "Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs" (arXiv 2309.05516), Intel, Findings of EMNLP 2024. REAP — "REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression" (arXiv 2510.13999), Cerebras Systems. Amanani e-perplexity asuka ku-AWQ (arXiv 2306.00978) nase-GPTQ (arXiv 2210.17323).

05I-Flash ngokumelene ne-Pro esiphethweni

Isigaba 03 sihlanganise izisekelo. Nakhu okubukeka kuwo womabili amazinga njengemikhiqizo esangweni lethu:

	DeepSeek-V4-Flash-0731	DeepSeek-V4-Pro-0813
Ukukalwa	W4A16-AUTOROUND-REAP	3BIT-REAP
Ukunemba kwezisindo	Amabhithi amane, ukugoba okufundiwe	Amabhithi amathathu
Isisekelo (esewonke / esebenzayo)	~304B / 13B	~1.7T / 49B
Okungenayo nge-1M yamathokheni	$0.50	$1.95
Okuphumayo nge-1M yamathokheni	$2.00	$3.90
Umongo / okukhishwayo okuphezulu	262,144 / 16,384	262,144 / 16,384
Amathuluzi · okukhishwayo okuhlelekile · ukusakaza	Yebo	Yebo
Umbono	Cha	Cha
Ungqimba lokwenqaba	Alukho	Alukho

Qaphela okungashintshi: umongo, uphahla lokukhishwayo, ukubiza amathuluzi, okukhishwayo okuhlelekile, kanye nokungabikho kongqimba lokwenqaba kuyafana. Izinto kuphela ezishintshayo yisisekelo, ukukalwa, kanye nentengo. Lokho kwenza ukushintsha phakathi kwazo kube ushintsho lomugqa owodwa ekhodini yakho nesinqumo esingesomnotho kuphela ekwakhiweni kwakho.

Kubiza malini ukubiza okukodwa kosayizi ophezulu

Gcwalisa wonke umongo wamathokheni angu-262,144 bese uthatha okukhishwayo okugcwele kwamathokheni angu-16,384:

Ngentengo yohlu. I-Flash: $0.131 engenayo + $0.033 ephumayo. I-Pro: $0.511 + $0.064. Isilinganiso singu-3.51×. I-Flash iphinde yabelane ngephansi lika-$0.50 / $2.00 ne-MiniMax-M3, i-Laguna-S-2.1, i-MiMo-V2.5, i-MiMo-V2.5-Pro ne-Hy3 — bheka ikhathalogu ephelele ukuze ubone ukuthi lezo zihluka kanjani ngombono nangokukhishwayo okuhlelekile.

06Izibalo zevolumu, zisetshenzisiwe

U-"eshibhe ngokuphindwe ka-3.9" uyiqiniso kokungenayo kuphela. Kokuphumayo i-Flash ishibhe ngokuphindwe ka-1.95. Ukonga kwakho kwangempela ngakho kuxhomeke esilinganisweni sakho sokungenayo nokuphumayo, futhi ukusabalala kubanzi ngokwanele ukushintsha izinqumo zokwakha. Imisebenzi emine engokoqobo, ngayinye ezicelweni ezingu-100,000:

Umsebenzi	Engenayo / ephumayo ekubizeni	Flash	Pro	Ukonga
Ukuhlukanisa amathikithi	1,200 / 120	$84	$281	3.3×
Ukuphendula imibuzo nge-RAG	8,000 / 600	$520	$1,794	3.4×
Isifinyezo sedokhumenti ende	60,000 / 1,500	$3,300	$12,285	3.7×
Ukukhiqizwa kwedatha yokwenziwa	500 / 2,000	$425	$878	2.1×

Iphethini: lapho usunduza umongo omningi futhi ukhiqiza umbhalo omncane, yilapho i-Flash ikonga khona kakhulu. Imishini yokuthola, ukuhlungwa kwamadokhumenti, ukuhlaziywa kwamalogi nokubhalwe phansi, kanye nokukhishwa kolwazi komongo omude yizona zindawo ezimnandi — zibuswa ngokufakwayo ngezinga eliphindwe kayishumi, futhi okungenayo yilapho isaphulelo esiphindwe ka-3.9 sihlala khona. Uma kunalokho umsebenzi wakho ungumyalelo omfushane okhiqiza amathokheni ayizinkulungwane ombhalo, umehluko uncipha ube cishe okuphindwe kabili futhi impikiswano yamakhono ye-Pro iba namandla ngokulinganayo.

Esinye isibalo esifanele ukugcinwa ekhanda: ku-$0.50 ngethokheni eyisigidi engenayo, lonke iwindi lamathokheni angu-262,144 libiza amasenti angu-13.1 ukuligcwalisa. Ukubhala imiyalelo yomongo omude kuyeka ukuba umbuzo wesabelomali ngaleyo ntengo, okwesinye isikhathi okusho ukuthi ungeqa ukwakha ungqimba lokuthola ngokuphelele bese uvumela iwindi lomongo lenze umsebenzi — ukonga kobunjiniyela ngaphezu kokonga kwamathokheni.

07Kusho ukuthini u-"ongavinjelwe" lapha, ngokuqondile

Igama lisetshenziswa ngokuxegayo, ngakho nayi incazelo yethu, ishiwo njengesakhi sesiphetho: asikho isihlukanisi sokwenqaba phambi kwemodeli futhi akukho ukuhlungwa kokuqukethwe okusetshenziswa kokukhishwayo kwayo. Akukho okubamba umyalelo bese kubuyisela ukwenqaba okulungiselelwe ngaphambili; akukho okuhlola ukugcwaliswa bese kukususa. Emsebenzini wokukhiqiza wevolumu ephezulu imiphumela ayishaqisi kangako futhi iwusizo ngaphezu kwalokho igama elikushoyo:

• Akukho ukwenqaba kwamanga kwithrafikhi engenabungozi. Iziphetho ezihlungiwe zivamise ukwenqaba izifinyezo zezokwelapha, imibhalo yokuphepha, ukutholwa kwezomthetho, izinganekwane, izinto zokunciphisa ingozi kanye nomsebenzi wokulawula okuqukethwe uqobo — ngqo lezo zigaba lapho umsebenzi weqoqo ucubungula umbhalo omayelana nesihloko esibucayi kunokukhiqiza noma yini elimazayo. Ezicelweni eziyizinkulungwane eziyikhulu, izinga lokwenqaba kwamanga elingu-1% liyimigqa eyinkulungwane ephukile.
• Izivumelwano ezicacile. Cela impendulo ye-json_schema bese uthola i-JSON ehambisana nesikimu, hhayi isigaba sokuxolisa lapho kufanele kube nento. Isihlaziyi sakho asidingi igatsha lokwenqaba, futhi umqondo wakho wokuzama kabusha akudingeki ukuthi ahlukanise ukwehluleka kokuthutha kokwenqubomgomo.
• Ungumnikazi wenqubomgomo. Yilokho ukushintshana. Uma umkhiqizo wakho udinga izivikelo zabasebenzisi bakho bokugcina, uzakha ezingeni lakho, lapho wazi khona abasebenzisi bakho nendawo yakho yomthetho, kunokuzuza ukuqagela komhlinzeki ngakho kokubili. I-akhawunti yakho ihlala ibuswa yiNqubomgomo Yethu Yokusebenzisa Ngokuzibophezela.

Okungavinjelwe akufani nokungaqondaniswa noma ukungasizi. Imodeli isalandela imiyalelo futhi isahlonipha umyalelo wesistimu. Uma ufuna ukuthi yenqabe izigaba ezithile kubasebenzisi bakho, kusho lapho — leyo miyalelo manje ingeyakho ukuyibhala kunokuba ngeyethu ukuyiphoqa.

08Amakhono nemikhawulo eqotho

Amakhono amabili aphethe ingxenye enkulu yomsebenzi wevolumu. Okukhishwayo okuqinile kwe-json_schema kuguqula ukugcwaliswa kube ngumugqa wesizindalwazi ngaphandle kongqimba lokuhlaziya nangaphandle kwegatsha lokwenqaba. Ukubiza amathuluzi kuzo zontathu izilimi kwenza i-Flash ibe yinjini enengqondo yezinyathelo ezincane ezishibhile ze-ejenti umhleli walo osezingeni eliphezulu osebenza kwimodeli enamandla kakhulu. Faka u-reasoning_effort, okukuvumela ukuthi uchithe amathokheni engeziwe kuphela kwingxenye yethrafikhi ewadingayo, kanye nomongo wamathokheni angu-262,144 ofana nowe-Pro — awulahli umongo ngokukhetha izinga elishibhile, okungajwayelekile.

Imikhawulo ibonakala ngokulinganayo:

• Awukho umbono. I-Flash ingeyombhalo kuphela futhi izokwenqaba izingxenye zesithombe emlayezweni. Ngezithombe kwimodeli engavinjelwe ephansini elifanayo lika-$0.50 / $2.00, i-MiniMax-M3 ne-MiMo-V2.5 yizo izinketho ezinombono; i-Kimi-K2.6 ne-Kimi-K3 zingamamodeli ombono anamandla kakhulu ngaphezu kwalelo zinga.
• Akusiyo eyaphezulu esikhwelweni. Amanani e-DeepSeek uqobo abeka i-Pro phambili kukho konke ukuhlolwa kwama-ejenti okushicilelwe, umehluko ubanzi kakhulu ekucabangeni kwekhodi okulandelanayo okunezinyathelo eziningi. Lapho impendulo eyodwa engalungile ibiza kakhulu kunezinkulungwane eziyodwa ezishibhile, chitha u-$1.45 owengeziwe ngesigidi ku-DeepSeek-V4-Pro.
• Amathokheni angu-16,384 okukhishwayo esicelweni ngasinye. Ukukhiqizwa okude kufanele kuhlukaniswe kubizwa kaningi. Lolu uphahla olufanayo kuyo yonke imodeli esiyikhonzayo, futhi luvala kudala ngaphambi kokuba iwindi lomongo lenze njalo.

09Ukuyibiza

Khomba noma yiliphi iklayenti elihambisana ne-OpenAI esangweni bese ushintsha uchungechunge lwemodeli. Akukho okunye esitakini sakho okufanele kuhambe.

curl https://api.shannon-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $SHANNON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP",
    "messages": [
      {"role": "system", "content": "Label the ticket. Reply only with the schema."},
      {"role": "user", "content": "The invoice PDF renders blank on Safari 18."}
    ],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "ticket_label",
        "schema": {
          "type": "object",
          "properties": {
            "category": {"type": "string"},
            "severity": {"type": "integer"},
            "needs_human": {"type": "boolean"}
          },
          "required": ["category", "severity", "needs_human"],
          "additionalProperties": false
        }
      }
    }
  }'

I-id yemodeli efanayo isebenza ingashintshiwe ku-/v1/messages naku-/v1/responses, okuyindlela esheshe kunazo zonke yokuhlola imodeli eshibhile ngaphansi kokuhlanganiswa okukhona ngaphandle kokubhala kabusha iklayenti lakho. Amapharamitha, izimo zezenzakalo zokusakaza, izikimu zokubiza amathuluzi kanye nemikhawulo yamazinga kubhalwe kureferensi ye-API; yonke imodeli ekhathalogu ibaliwe ku-Shannon Research.

10Imibuzo ebuzwa njalo

---

Imithombo: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · DeepSeek-V3 Technical Report (arXiv 2412.19437) · DeepSeek-V4 (arXiv 2606.19348) · amakhadi emodeli e-DeepSeek-V4-Flash-0731 nele-DeepSeek-V4-Pro-0813 · AutoRound (arXiv 2309.05516, Intel) · REAP (arXiv 2510.13999, Cerebras Systems) · AWQ (arXiv 2306.00978) · GPTQ (arXiv 2210.17323). Wonke amanani esisekelo nawokuhlolwa anjengoba eshicilelwe ngababhali bawo; amanani e-Shannon AI nomongo osetshenziswayo angamanani ethu ohlu kusukela ngoSepthemba 2026.