DeepSeek Flash API Yosasidwa

Magulu ambiri amene akuyesa API ya LLM amakonza chinthu cholakwika. Amayeza pa mafunso makumi awiri ovuta kwambiri amene angaganize, amasankha wopambana, ndipo pambuyo pa milungu isanu ndi umodzi amazindikira kuti makumi asanu ndi asanu ndi limodzi peresenti ya magalimoto awo ndi ntchito yaifupi, yokonzedwa, yobwerezabwereza imene mtundu wolipiritsa gawo limodzi mwa anayi ukanachita chimodzimodzi. DeepSeek-V4-Flash ilipo chifukwa cha peresenti khumi ndi zisanu zimenezo. Tsambali likufotokoza za kuzindikira pamene muli mu gawo limenelo — ndi pamene simuli.

01Chomwe mumapeza kwenikweni

Chinthu	DeepSeek-V4-Flash-0731 · W4A16-AUTOROUND-REAP
Id ya mtundu	DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP
Context / kutuluka kwambiri	ma token 262,144 / ma token 16,384
Mtengo pa 1M (kulowa / kutuluka)	$0.50 / $2.00
Streaming · kuyitanitsa ntchito · kuganiza	Inde — zonse zitatu
Zotuluka zokonzedwa	json_schema ndi response_format
Zolowetsa za zithunzi (masomphenya)	Ayi — mawu okha
Wosanjikiza wa kukana / kusefa kwa zotulutsa	Palibe
Yotumikiridwa pa	Gulu la Shannon la ma GPU lokha

Zilankhulo zitatu za API zimafikira zolemera zomwezo: /v1/chat/completions (mawonekedwe a OpenAI), /v1/messages (mawonekedwe a Anthropic), ndi /v1/responses. Streaming imagwira ntchito pa zonse zitatu. Mndandanda wonse wa ma parameter uli mu zolemba za API.

02Komwe amachokera: mbadwo wa DeepSeek

DeepSeek-AI yakhala ndi mibadwo isanu ya mitundu ikuyang'ana chinthu chimodzi: kupeza machitidwe apamwamba kuchokera ku mtundu wotsika mtengo ku kutumikira, osati wotsika mtengo kuphunzitsa kokha. Pafupifupi chisankho chilichonse cha kapangidwe m'banjali chimachokera pamenepo, ndipo zofunika kwambiri zafalitsidwa. Zitatu zimafotokoza chomwe mukuyitana pamene mwafika pa endpoint iyi.

DeepSeekMoE — akatswiri ang'onoang'ono kuphatikiza wina wogawana

Chosanjikiza cha Mixture-of-Experts chimasintha netiweki imodzi yayikulu ya feed-forward ndi zambiri zazing'ono kuphatikiza router imene imasankha zochepa pa token iliyonse. Mtundu wa DeepSeek umadula akatswiri ang'onoang'ono kwambiri kuposa mawonekedwe wamba a akatswiri asanu ndi atatu, choncho ma FLOP omwewo ogwiritsidwa ntchito amagula kusakaniza kolondola kwambiri, ndipo umapatula katswiri wina wogawana amene token iliyonse imadutsa mwa iye, choncho chidziwitso chofala sichiphunziridwanso mobwerezabwereza mkati mwa katswiri aliyense. DeepSeek-V3 inatulutsa izi ngati akatswiri 256 ogawidwa ntchito kuphatikiza katswiri 1 wogawana, kuyendetsa 8 ogawidwa ntchito pa token iliyonse — zigawo zonse 671B, 37B zogwira ntchito — kuphatikiza kulinganiza katundu wopanda kutayika kwa auxiliary-loss: m'malo mwa mtengo wolangira umene umachepetsa mtundu, chikoka cha katswiri aliyense chimakankhira zisonkhezero za router pamwamba kapena pansi sitepe iliyonse potengera ngati katswiriyo akugwira ntchito zochepa kapena zambiri.

MLA — chisamalira chimene chimakwana mu kukumbukira

Theka lina la vuto la mtengo wotumikira ndi KV cache, imene imakula mofanana ndi context ndipo imalamulira kukumbukira pa kutalika kwakukulu kwa mizere. Multi-head Latent Attention, yoyambitsidwa mu DeepSeek-V2, imapanikiza makiyi ndi mitengo pamodzi kukhala vekitala ya latent yotsika-rank ndipo imasunga chimenecho, kumanganso K/V yogwiritsika nthawi yomweyo. Ziwerengero za DeepSeek zofalitsidwa za V2, poyerekeza ndi DeepSeek 67B yawo yakale yolimba, zikunena za kuchepetsa kwa 93.3% kwa KV cache ndi kutulutsa kwakukulu kwa 5.76× kuposa kale. Ndicho chifukwa chimene zenera la ma token 262K limakhala chizolowezi m'banjali m'malo mokhala chowonjezera chapamwamba.

Chisamalira chosakanikirana — ndi chitsanzo cha ma token otsika mtengo

V3.2-Exp inawonjezera DeepSeek Sparse Attention: chowunikira chopepuka chimayerekezera ma token a context ndi funso, ndipo chosankha chimasunga zochepa zomwe zayerekezedwa kwambiri. Chofunika kwambiri ndi umboni, osati njira yokha. DeepSeek inafalitsa kuyerekezera pambali ndi V3.1-Terminus ndi MMLU-Pro yofanana pa 85.0, AIME 2025 yokwera (88.4 → 89.3), SWE-bench Verified yotsika pang'ono (68.4 → 67.8) — mtundu wofanana pafupifupi — ndipo anadula mitengo yawo ya API ndi kupitirira 50% tsiku lomwelo. Kapangidwe, kofalitsidwa moona, kukusintha mwachindunji kukhala mtengo. Flash ndicho chinenero chimenecho chotengedwa sitepe imodzi kupitirira.

Mbadwo wa V4

V4, yowonetsedwa mu April 2026 pansi pa layisensi ya MIT, imasintha unyolo wa chisamalira kachiwiri ndi kusakanizidwa kwa Compressed Sparse Attention ndi Heavily Compressed Attention, imaphunzitsidwa ndi optimizer ya Muon pa ma token oposa 32T, ndipo imasunga cholinga cha V3 cha multi-token-prediction — chimene chimagwiranso ntchito ngati speculative decoding, ndipo DeepSeek anayeza kuti chimakhala ndi mlingo wa kuvomereza token yachiwiri wa 85–90% pofulumizitsa decoding pafupifupi 1.8× mu V3. Chidule cha V4 chikunena kuti V4-Pro imayenda pa 27% ya FLOP za V3.2 za single-token inference ndi 10% ya KV cache yake. Banjali limasonyeza milingo itatu ya reasoning-effort, ndicho chimene parameter yathu ya reasoning_effort imayimira.

Magwero: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · DeepSeek-V3 Technical Report (arXiv 2412.19437) · DeepSeek-V3.2-Exp model card · DeepSeek-V4 (arXiv 2606.19348). Ziwerengero zonse zafalitsidwa ndi DeepSeek.

03Chomwe "Flash" amatanthauza kwenikweni

Ichi ndi malo amene poyerekeza ambiri amalakwitsa, choncho ndikofunika kunena momveka: Flash ya DeepSeek si Pro yolembedwanso ndi quantization, kudulidwa kapena distillation. Ndi mtundu wophunzitsidwa padera, wang'onoono wa Mixture-of-Experts wokhala ndi kuphunzitsa kwake kokha. Pepala la V4 preview likupereka makulidwe awiriwo:

Ziwerengero zochokera ku pepala la DeepSeek-V4 preview. Ma checkpoint enieni amene timatumikira — V4-Flash-0731 ndi V4-Pro-0813 — amasonyeza zigawo zonse zazikulu pang'ono za 304B ndi 1.7T pa makaadi awo a mtundu; DeepSeek satchulanso zigawo zogwira ntchito za ma build amenewo.

Ziwerengero ziwiri ndizo zofunika pano. Zigawo zonse zimaika malo a kukumbukira, zimene zimatsimikizira kuti kope limodzi la mtunduwo limagwiritsa ntchito zida zingati. Zigawo zogwira ntchito zimaika masamu pa token iliyonse, zimene zimatsimikizira kutulutsa. Flash ndi yaying'ono pafupifupi kuwirikiza 5.6× pa choyamba ndi kuwirikiza 3.8× pa chachiwiri — ndipo pali chinthu chodabwitsa chowoneka mu kusiyana kumeneko: Flash imagwiritsa ntchito 13 mwa zigawo 284 biliyoni, pafupifupi 4.6%, pomwe Pro imagwiritsa ntchito 49 mwa 1,600 biliyoni, pafupifupi 3.1%. Flash ndi mtundu wosakanikirana pang'ono. Si Pro yokhala ndi akatswiri ochepa oyatsidwa; ndi kapangidwe kolimba kwambiri, kolinganizidwa mosiyana, kamene kamagwiritsa ntchito gawo lalikulu la kake pa token iliyonse imene kamatulutsa. Ichi ndi chisankho chenicheni cha uinjiniya, osati kutsikira pansi.

Zomwe ziwerengero za DeepSeek zomwe zikunena

Makaadi onse aakulu a mtundu amafalitsa zotsatira pa ma benchmark omwewo a agentic ndi software-engineering, zomwe zimapangitsa kuyerekezera kumeneku kukhala koyera modabwitsa — labotale yomweyo, njira yomweyo yoyeserera, makulidwe awiri:

Benchmark (yofalitsidwa ndi 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

Pro imapambana mzere uliwonse. Ichocho ndi mutu weniweni, ndipo tikanakonda kuti muwerenge pano m'malo mokuzindikira mukugwira ntchito. Koma werengani malire, chifukwa ndicho chimene mukugula kwenikweni. Pa Terminal Bench 2.1 mpata ndi mfundo 5.2 pa mtundu wokhala ndi gawo limodzi mwa makumi awiri ndi zisanu la zigawo. Pa kugwiritsa ntchito zida ndi mfundo 3.8. Mpatawo umakulirakulira pamene ntchito ikuya kwambiri — mfundo 8.3 pa DeepSWE, 7.3 pa NL2Repo, zonsezi zimapatsa mphotho kuganiza kwa masitepe ambiri motsatizana pa code. Icho ndi mawonekedwe a kusinthana konseko mu tebulo limodzi: Flash imatseka mtunda waukulu pa ntchito zamalire ndipo imasiya malo pa zosatha.

Kaadi yeniyeni ya Flash ya DeepSeek ikutchula mfundoyo momveka bwino kwambiri: Flash-0731 imapambana V4-Pro preview pa ma benchmark otchulidwa ngakhale ili ndi zigawo zogwira ntchito zochepa kwambiri. Mtundu waung'ono watsopano kupambana mtundu waukulu wakale ndi mkhalidwe wamba m'munda uwu, ndipo ndicho chifukwa chake "mlingo waukulu" sungakhale yankho lanu lokhazikika.

04Kuwerenga chowonjezera: W4A16-AUTOROUND-REAP

Ma id athu a mitundu amanena momveka njira yawo ya kupanikiza, chifukwa ndicho chinthu chimene chimakhudza kwambiri mtundu wa zotulutsa ndipo pafupifupi palibe amene amafotokoza. Chimenechi chimagawika m'magawo atatu.

W4A16 — zolemera za 4-bit, zochitika za 16-bit

Zolemera zimasungidwa pa mabiti 4 zilizonse; zochitika zimayenda pa kulondola kwa 16-bit. Ichi ndi quantization ya zolemera zokha, ndipo imayang'ana vuto lenileni la kutulutsa ma token, limene ndi kufulumira kwa memory bandwidth m'malo mwa masamu. Kuika quantization pa zochitika nazonso kukanagula kufulumira pang'ono ndipo kukanapereka ngozi yayikulu ya kulondola; kusiya pa 16-bit ndicho chisankho chosamala, cholondola kwambiri. Mawuwa ndi njira yachidule ya dziko la kutumikira — mabuku a quantization amalemba chomwecho ngati W4 wokhala ndi kukula kwa gulu.

AUTOROUND — kuzungulira kophunziridwa, osati kuzungulira ku pafupi

Quantization yosaganizira imazungulira kulemera kulikonse ku mtengo woyandikira kwambiri wowonetsedwa. Izi zaonetsedwa kuti si zabwino kwambiri: kuzungulira kulemera kumodzi "mwachilakwa" kungathe kufafaniza cholakwika chochokera ku zoyandikana nazo, ndipo pa mabiliyoni a zolemera zolumikizana zimenezo zimalamulira. AutoRound, yochokera ku Intel, imalingalira chisankho cha kuzungulira chokha ngati chinthu choti chiphunziridwe — kusintha kophunziridwa pa kulemera kulikonse kokakamizidwa pa ±0.5, kokwanira kusintha chisankho chilichonse cha kuzungulira ndipo osaposa apo, kuphatikiza malire ophunziridwa a kudula, okonzedwa ndi signed gradient descent motsutsana ndi cholinga cha kumangidwanso kwa zotulutsa pa gulu, pa gulu laling'ono la kuyeza. Masitepe mazana awiri, ndiye zatha.

Kuyerekezera kofalitsidwa ndi Intel kukusonyeza AutoRound ikupambana GPTQ mu 30 mwa zoyeza 32, AWQ mu 27 mwa 32 ndi OmniQuant mu 29 mwa 32 pa makonzedwe a zolemera zokha za 4-bit, ndi kupindula pa kuzungulira ku pafupi mosaganizira kwa mfundo 6.9 mpaka 33.2 pa mabiti 2. Palibe pa zonsezi chimene chimalipiritsa chilichonse pa nthawi ya kuganiza: ndi chisankho chabwino kwambiri cha manambala, chopangidwa kamodzi, kunja kwa nthawi ya ntchito.

REAP — kudula akatswiri amene palibe amene amawapatsa ntchito

REAP (Router-weighted Expert Activation Pruning), yochokera ku Cerebras Systems ndi University of Calgary, ndi kupanikiza kamodzi kokha ka mitundu ya MoE. Imayerekezera katswiri aliyense potengera kulimba ndi kuchuluka kwa nthawi imene router imamugawira ntchito, kuwirikizidwa ndi kukula kwa zotulutsa za katswiriyo, kenako imachotsa akatswiri ochepa kwambiri ofunika — popanda kuphunzitsanso pambuyo pake.

Chopezeka chachikulu cha pepalali ndi chogwiritsika: pa magawo apamwamba a kupanikiza, kudula akatswiri kumapambana kuphatikiza kwawo. Kuphatikiza akatswiri osiyana pa ntchito kumayambitsa chimene olemba amatcha functional subspace collapse — router imataya ulamuliro wolondola, wodalira zolowetsa umene umapangitsa MoE kukhala yofunika kumangidwa, ndipo imasiyidwa kuyendetsa pakati pa avereji zosasintha. Ziwerengero zawo pa mitundu yochokera ku 20B mpaka 1T zigawo: pa kupanikiza kwa akatswiri kwa 25% pa ntchito za code, REAP imataya 2.8% ya kulondola kwa avereji pomwe njira zophatikiza zimataya kupitirira 5%; pa 50%, REAP imataya 8.0% pomwe kuphatikiza kumataya kupitirira 20%. Pa Qwen3-Coder-480B pa kudula kwa 50% kutsika ndi 1.4%, motsutsana ndi 39.3% pa kudula kosaganizira kozikidwa pa kuchuluka. Cerebras yafalitsa ma checkpoint a REAP a mzere wa DeepSeek pakati pa ena.

Chifukwa cha 4-bit apa ndi 3-bit pa Pro

Milingo yonse iwiri iyenera kukwana ndi kupitirira kufulumira pa gulu lomwelo la zida. Pro imanyamula matiriliyoni 1.7 a zigawo, choncho imafunika chithandizo cholimba kwambiri cha 3-bit kuti ikwane. Flash imanyamula pafupifupi mabiliyoni 300, choncho sifunika — ndipo biti imene imasunga imagwiritsidwa ntchito pa kulondola m'malo mokwana.

Biti yowonjezerayo imakhudza kupambana kuposa masamu ake amene amawonetsa, chifukwa kuwonongeka kwa quantization sikofanana. Mu miyeso ya WikiText-2 ya pepala la AWQ, Llama-2-7B imachoka ku perplexity ya 5.47 pa FP16 kupita ku 5.60 pa 4-bit — chilango cha 2.4% — koma kupita ku 6.24 pa 3-bit, chilango cha 14%. Imeneyo ndi pafupifupi kuwirikiza kasanu ndi kamodzi kwa kuwonongeka pa 25% ya kupanikiza kowonjezera, ndipo mawonekedwe omwewo amagwira ntchito pa 70B. Pepala la GPTQ ndi lowonjezera momveka: pa mabiti 3, kuzungulira ku pafupi mosaganizira kumagwa kwathunthu pa mitundu ya gulu la 175B, kufika pa masauzande a perplexity, pomwe kukhalabe kogwiritsika bwino pa 4-bit. Pansi pa mabiti anayi, algorithm yabwino ya quantization imasiya kukhala chowonjezera ndipo imakhala chinthu chokha chimene chikusunga mtunduwo pamodzi.

Choncho kuwerenga koona kwa zowonjezera ziwirizi ndi: Flash imakhala pafupi kwambiri ndi ubwino wake wosapanikizidwa kuposa Pro imene imakhala pafupi ndi wake. Pro ikadali mtundu wamphamvu kwambiri — tebulo la benchmark pamwambapa likutsimikiza izi. Kokha inapereka zambiri kuti ifike pano, ndipo mumalipira mwayi umenewo.

Magwero: 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. Ziwerengero za perplexity zochokera ku AWQ (arXiv 2306.00978) ndi GPTQ (arXiv 2210.17323).

05Flash motsutsana ndi Pro pa endpoint

Gawo 03 lidafotokoza za maziko. Nayi mmene milingo iwiriyo imawonekera ngati zinthu pa gateway yathu:

	DeepSeek-V4-Flash-0731	DeepSeek-V4-Pro-0813
Quantization	W4A16-AUTOROUND-REAP	3BIT-REAP
Kulondola kwa zolemera	4-bit, kuzungulira kophunziridwa	3-bit
Maziko (zonse / zogwira ntchito)	~304B / 13B	~1.7T / 49B
Kulowa pa 1M ya ma token	$0.50	$1.95
Kutuluka pa 1M ya ma token	$2.00	$3.90
Context / kutuluka kwambiri	262,144 / 16,384	262,144 / 16,384
Zida · zotuluka zokonzedwa · streaming	Inde	Inde
Masomphenya	Ayi	Ayi
Wosanjikiza wa kukana	Palibe	Palibe

Onani chomwe sichisintha: context, malire a kutuluka, kuyitanitsa ntchito, zotuluka zokonzedwa, ndi kusapezeka kwa wosanjikiza wa kukana ndizofanana. Zinthu zokha zomwe zimasintha ndi maziko, quantization, ndi mtengo. Izi zimapangitsa kusinthira pakati pawo kukhala kusintha kwa mzere umodzi wa mawu mu code yanu ndi chisankho cha zachuma chokha mu kapangidwe kanu.

Chomwe mayitanidwe amodzi akulu kwambiri amalipiritsa

Dzazani context yonse ya ma token 262,144 ndipo mutenge kutuluka konse kwa ma token 16,384:

Pa mtengo wa mndandanda. Flash: $0.131 kulowa + $0.033 kutuluka. Pro: $0.511 + $0.064. Chiwerengero 3.51×. Flash imagawananso mtengo wotsika kwambiri wa $0.50 / $2.00 ndi MiniMax-M3, Laguna-S-2.1, MiMo-V2.5, MiMo-V2.5-Pro ndi Hy3 — onani mndandanda wonse kuti mudziwe momwe zimasiyana pa masomphenya ndi zotuluka zokonzedwa.

06Masamu a kuchuluka, owerengedwa

"Kuwirikiza 3.9× kutsika mtengo" ndikoona pa kulowa kokha. Pa kutuluka Flash ndi yotsika kuwirikiza 1.95×. Kupindula kwanu kwenikweni choncho kumadalira chiŵerengero cha kulowa-ndi-kutuluka kwanu, ndipo kusiyana ndi kwakukulu mokwanira kusintha zisankho za kapangidwe. Ntchito zenizeni zinayi, iliyonse pa mayitanidwe 100,000:

Ntchito	Kulowa / kutuluka pa yitanidwe	Flash	Pro	Kupindula
Kuika m'magulu ma tikiti	1,200 / 120	$84	$281	3.3×
Kuyankha mafunso a RAG	8,000 / 600	$520	$1,794	3.4×
Chidule cha chikalata chachitali	60,000 / 1,500	$3,300	$12,285	3.7×
Kupanga deta yopeka	500 / 2,000	$425	$878	2.1×

Mawonekedwe ake: mmene mumakankhira context yambiri ndi kutulutsa mawu ochepa, ndi mmene Flash imakupulumutsirani zambiri. Njira za kubwezeretsa, kusankha zikalata, kusanthula ma log ndi zolembedwa, ndi kutulutsa deta ya context yaitali ndizo malo abwino kwambiri — zimadalira context kwambiri kuposa kutuluka, ndipo kulowa ndi komwe chidule cha 3.9× chilipo. Ngati m'malo mwake ntchito yanu ndi pempho laifupi lopanga masauzande a ma token a mawu, mpata umachepa mpaka pafupifupi 2× ndipo mfundo ya kuthekera ya Pro imalimba mofanana.

Masamu ena ofunika kukumbukira: pa $0.50 pa miliyoni ya ma token olowa, zenera lonse la ma token 262,144 limalipiritsa masenti 13.1 kudzaza. Kufunsa kwa context yaitali kumasiya kukhala funso la ndalama pa mtengo umenewo, zomwe nthawi zina zikutanthauza kuti mungasiye kumanga chosanjikiza cha kubwezeretsa konse ndikulola zenera la context kuchita ntchitoyo — kupindula kwa uinjiniya pamwamba pa kupindula kwa ma token.

07Chomwe "chosasidwa" chimatanthauza pano, molondola

Liwuli limagwiritsidwa ntchito mosamala pang'ono, choncho nayi tanthauzo lathu, lofotokozedwa ngati mkhalidwe wa endpoint: palibe chowunika cha kukana chimene chili patsogolo pa mtunduwo ndipo palibe kusefa kwa zinthu kumene kumagwiritsidwa ntchito pa zotulutsa zake. Palibe chomwe chimayima pempho ndi kubweza kukana kokonzedweratu; palibe chomwe chimayang'ana zotulutsa ndi kuzichotsa. Pa ntchito za kuchuluka kwakukulu zotsatira zake ndi zochepa modabwitsa ndipo zothandiza kuposa mmene liwulo limawonetsera:

• Palibe zolakwika zabodza pa magalimoto abwino. Ma endpoint osefedwa nthawi zambiri amakana zidule za zachipatala, zolemba za chitetezo, kufufuza kwa milandu, nkhani zopeka, zolemba zochepetsera zowononga ndi ntchito ya kuwunika yokha — ndendende magulu amene ntchito ya bulki ikusanthula mawu okhudza mutu wovuta m'malo mopanga chinthu choopsa. Pa mayitanidwe 100,000, chiŵerengero cha kukana chabodza cha peresenti imodzi ndi mizere chikwi chowonongeka.
• Mapangano otsimikizika. Pemphani yankho la json_schema ndipo mumapeza JSON yogwirizana ndi schema, osati ndime yopepesa pomwe payenera kukhala chinthu. Parser yanu safuna nthambi ya kukana, ndipo njira yanu ya kuyesanso safunika kusiyanitsa kulephera kwa mayendedwe ndi kwa ndondomeko.
• Muli ndi ndondomeko yanu. Icho ndicho chosinthana. Ngati chinthu chanu chimafunikira zoteteza kwa ogwiritsa ntchito anu, mumazimanga pa mlingo wanu, kumene mumadziwa ogwiritsa ntchito anu ndi malamulo anu, m'malo molandira zoyerekeza za wopereka onse awiri. Akaunti yanu ikupitirira kuyendetsedwa ndi Ndondomeko yathu ya Kagwiritsidwe Kaudindo.

Chosasidwa sichofanana ndi chosagwirizana kapena chosathandiza. Mtunduwo ukupitirirabe kutsatira malangizo ndipo ukupitirirabe kulemekeza funso la dongosolo. Ngati mukufuna kuti ukane magulu ena kwa ogwiritsa ntchito anu, nenani izo pamenepo — lamuloli tsopano ndi lanu kulemba m'malo mwa lathu kukukakamizani.

08Maluso ndi malire enieni

Maluso awiri amanyamula gawo lalikulu la ntchito ya kuchuluka. Zotuluka za json_schema zolimba zimasandutsa zotulutsa kukhala mzere wa database popanda chosanjikiza cha kusanthula ndi popanda nthambi ya kukana. Kuyitanitsa ntchito mu zilankhulo zonse zitatu zimapangitsa Flash kukhala injini yomveka bwino ya masitepe otsika mtengo a agent imene woyendetsa wamkulu umayenda pa mtundu wamphamvu kwambiri. Wonjezerani reasoning_effort, imene imakulolani kugwiritsa ntchito ma token owonjezera pa gawo lokha la magalimoto amene amawafunikira, ndi context ya ma token 262,144 yofanana ndi ya Pro — simupereka context poswanka mlingo wotsika mtengo, chomwe ndi chinthu chosazolowereka.

Malire ndi enieni mofanana:

• Palibe masomphenya. Flash ndi mawu okha ndipo idzakana magawo a zithunzi mu uthenga. Pa zithunzi mu mtundu wosasidwa pa mtengo womwewo wotsika kwambiri wa $0.50 / $2.00, MiniMax-M3 ndi MiMo-V2.5 ndizo zosankha zomwe zimalandira masomphenya; Kimi-K2.6 ndi Kimi-K3 ndi mitundu yamphamvu kwambiri ya masomphenya pamwamba pa mlingowo.
• Si pamwamba pa makwerero. Ziwerengero za DeepSeek zokha zikuyika Pro patsogolo pa benchmark iliyonse yofalitsidwa ya agentic, ndi mpata waukulu kwambiri pa kuganiza kwa masitepe ambiri motsatizana pa code. Kumene yankho limodzi lolakwika limalipiritsa kuposa chikwi cha otsika mtengo, gwiritsani ntchito $1.45 yowonjezera pa miliyoni pa DeepSeek-V4-Pro.
• Ma token 16,384 otuluka pa yitanidwe. Zotulutsa zazitali ziyenera kugawidwa m'mayitanidwe angapo. Ili ndi malire omwewo pa mtundu uliwonse umene timatumikira, ndipo umagwira ntchito kale kuposa zenera la context.

09Kuyitana

Ikani kasitomala aliyense wogwirizana ndi OpenAI pa gateway ndipo sinthani mawu a mtundu. Palibe china chilichonse mu kapangidwe kanu chomwe chikuyenera kusuntha.

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
        }
      }
    }
  }'

Id yomweyo ya mtundu imagwira ntchito popanda kusintha pa /v1/messages ndi /v1/responses, njira yofulumira kwambiri yoyesera A/B mtundu wotsika mtengo pansi pa kaphatikizidwe kamene kalipo popanda kulembanso kasitomala wanu. Ma parameter, mawonekedwe a zochitika za streaming, ma schema a kuyitanitsa ntchito ndi malire a kuchuluka ndi zolembedwa mu malongosoledwe a API; mtundu uliwonse mu mndandanda uli pa Kafukufuku wa Shannon.

10Mafunso ofunsidwa kawirikawiri

---

Magwero: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · DeepSeek-V3 Technical Report (arXiv 2412.19437) · DeepSeek-V4 (arXiv 2606.19348) · makaadi a mtundu a DeepSeek-V4-Flash-0731 ndi DeepSeek-V4-Pro-0813 · AutoRound (arXiv 2309.05516, Intel) · REAP (arXiv 2510.13999, Cerebras Systems) · AWQ (arXiv 2306.00978) · GPTQ (arXiv 2210.17323). Ziwerengero zonse za maziko ndi benchmark ndi zofalitsidwa ndi olemba awo; mitengo ndi context yotumikiridwa ya Shannon AI ndi zathu zokha za mndandanda pofika Seputembala 2026.