API yeDeepSeek Flash Isina Kudzorwa

Mapoka mazhinji anoongorora API yeLLM anogadzirisa chinhu chisiriicho. Anoedza pamaprompt makumi maviri akaoma kupfuura ose aanogona kufunga, osarudza inokunda, ozoona pashure pemavhiki matanhatu kuti makumi mapfumbamwe nemashanu muzana ehuwandu hwawo ibasa pfupi, rakarongwa, rinodzokororwa raizoitwa nemodhi inodhura chikamu chechina zvakafanana chaizvo. DeepSeek-V4-Flash iripo nekuda kwemakumi mapfumbamwe nemashanu muzana iwayo. Peji rino nderekuziva kana uri mairi — uye kana usiri.

01Chaunowana chaizvo

Chiratidzo	DeepSeek-V4-Flash-0731 · W4A16-AUTOROUND-REAP
Zita remodhi	DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP
Context / zvinobuda zvakanyanya	matokeni 262,144 / matokeni 16,384
Mutengo pa1M (anopinda / anobuda)	$0.50 / $2.00
Kuyerera · kudana zvishandiso · kufungisisa	Hongu — mitauro mitatu yose
Zvinobuda zvakarongwa	json_schema neresponse_format
Mifananidzo inopinda (kuona)	Kwete — mavara chete
Chidzitiro chekuramba / sefa pane zvinobuda	Hapana
Inoshandiswa pa	cluster yedu yemaGPU

Mitauro mitatu yeAPI inosvika pahuremu humwe chete: /v1/chat/completions (chimiro cheOpenAI), /v1/messages (chimiro cheAnthropic), ne/v1/responses. Kuyerera kunoshanda pamitatu yose. Rondedzero yakazara yemapharamita iri mumagwaro eAPI.

02Kwainobva: dzinza reDeepSeek

DeepSeek-AI yapedza zvizvarwa zvishanu zvemamodhi pachinhu chimwe chete: kuwana maitiro epamberi kubva pamodhi yakachipa kushandisa, kwete inongochipa kudzidzisa. Anenge sarudzo yega yega yemagadzirirwo mumutsara uyu inodzokera ipapo, uye dzinokosha dzakabudiswa. Zvitatu zvinotsanangura zvaunodana kana wasvika pasuwo rino.

DeepSeekMoE — nyanzvi diki-diki pamwe neimwe inogovaniswa

Danho reMixture-of-Experts rinotsiva netiweki imwe huru yefeed-forward nedzakawanda diki nerouter inosarudza shoma patokeni imwe neimwe. Rudzi rweDeepSeek runocheka nyanzvi diki-diki kupfuura chimiro chakajairika chenyanzvi sere, saka maFLOPs mamwe chetewo anoshanda anotenga musanganiswa wakanyatsotarwa zvikuru, uye rwunotsaura nyanzvi inogovaniswa inopfuurwa netokeni yega yega, saka ruzivo rwakajairika harudzokororwi ruchidzidzwazve mukati menyanzvi imwe neimwe. DeepSeek-V3 yakauya neizvozvo senyanzvi 256 dzinotumirwa pamwe nenyanzvi 1 inogovaniswa, ichishandisa 8 dzinotumirwa patokeni imwe neimwe — mapharamita 671B ose, 37B anoshanda — pamwe nekuenzanisa mutoro kusina auxiliary loss: panzvimbo pechirango chinodzikisira mhando, bhaizi panyanzvi imwe neimwe rinosundidzira zviyero zvekutumira mudenga kana pasi padanho rega rega sezvo nyanzvi iyoyo ichishanda zvishoma kana zvakanyanya.

MLA — attention inokwana mumemori

Rimwe divi redambudziko remutengo wekushandisa iKV cache, inokura zvakaenzana necontext uye inotonga memori pamitsara mirefu. Multi-head Latent Attention, yakaunzwa muDeepSeek-V2, inodzvanya makiyi nezvikoshwa pamwe chete kuita vekita yepasi ye latent yochengeta iyoyo, ichivakazve K/V inoshanda panguva yacho. Nhamba dzeV2 dzakabudiswa neDeepSeek, kuenzaniswa neDeepSeek 67B yavo yekare yakazara, dzinoshuma kudzikira kwe93.3% kweKV cache nekubudirira kwekugadzira kwakanyanya kunopetwa ka5.76. Ndicho chikonzero hwindo rematokeni 262K riri chinhu chakajairika mumhuri iyi pane chinhu chinowedzerwa pamutengo wepamusoro.

Sparse attention — nemuenzaniso wematokeni akachipa

V3.2-Exp yakawedzera DeepSeek Sparse Attention: indekisi yakareruka inopa matokeni econtext zviyero ichienzanisa nemubvunzo, uye musarudzi anochengeta ane zviyero zvepamusoro chete. Chinhu chinonakidza igwaro rekubhadhara, kwete nzira yacho. DeepSeek yakaburitsa kuenzanisa neV3.1-Terminus paine MMLU-Pro yakafanana pa85.0, AIME 2025 yakwira (88.4 → 89.3), SWE-bench Verified yakadzika zvishoma (68.4 → 67.8) — mhando yakati sandara — uye yakadzikisa mitengo yayo yeAPI nekupfuura 50% pazuva rimwe chetero. Magadzirirwo, akashumwa nechokwadi, achishandurwa zvakananga kuita mutengo. Flash inharo idzodzo dzakaendeswa danho rimwe mberi.

Chizvarwa cheV4

V4, yakaratidzwa muna Kubvumbi 2026 pasi perezinesi reMIT, inotsiva chimiro cheattention zvakare nemusanganiswa weCompressed Sparse Attention neHeavily Compressed Attention, inodzidzisa neMuon optimizer pamatokeni anopfuura 32T, uye inochengeta chinangwa cheV3 chekufanotaura matokeni akawanda — chinoshandawo se speculative decoding, uye icho DeepSeek yakayera pachiyero chekugamuchira tokeni yechipiri che85–90% chekumhanya kwedecoding kunopetwa kanenge ka1.8 muV3. Pfupiso yeV4 inoshuma V4-Pro ichimhanya pa27% yemaFLOPs eV3.2 etokeni imwe chete ne10% yeKV cache yayo. Mhuri inoburitsa mazinga matatu ekufungisisa, ndizvo zvinonongedzerwa nepharamita yedu yereasoning_effort.

Zvakatorwa kubva: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · Gwaro reunyanzvi reDeepSeek-V3 (arXiv 2412.19437) · kadhi remodhi reDeepSeek-V3.2-Exp · DeepSeek-V4 (arXiv 2606.19348). Nhamba dzose dzakashumwa neDeepSeek.

03Kuti "Flash" inorevei chaizvo

Apa ndipo panokanganisa kuenzanisa kwakawanda, saka zvakakodzera kutaurwa pachena: Flash yeDeepSeek haisi Pro yakaitwa quantization, yakabviswa nyanzvi kana yakadzidziswa kuitevedzera. Imodhi yeMixture-of-Experts yakadzidziswa yakaparadzana, diki, ine kudzidziswa kwayo kwekutanga. Gwaro rekuratidzwa kweV4 rinopa hukuru huviri:

Nhamba dzinobva mugwaro rekuratidzwa kweDeepSeek-V4. Machekipoinzi epamutemo atinoshandisa — V4-Flash-0731 neV4-Pro-0813 — anonyora huwandu hukuru zvishoma hwe304B ne1.7T pamakadhi awo emamodhi; DeepSeek haitaurizve mapharamita anoshanda pazvivakwa izvozvo.

Nhamba mbiri ndidzo dzinoita basa pano. Mapharamita ose anotara nzvimbo yememori, inosarudza kuti kopi imwe yemodhi inotora nhumbi dzakawanda sei. Mapharamita anoshanda anotara masvomhu patokeni, iyo inosarudza kubudirira. Flash idiki kanenge ka5.6 pane yekutanga uye ka3.8 pane yechipiri — uye pane chinhu chakadzama chakakodzera kuonekwa mumukaha iwoyo: Flash inoshandisa 13 kubva pamabhiriyoni 284, anenge 4.6%, apo Pro inoshandisa 49 kubva pamabhiriyoni 1,600, anenge 3.1%. Flash ndiyo modhi isina kupararira zvakanyanya. Haisi kungova Pro ine nyanzvi shoma dzakabatidzwa; imagadzirirwo akaunganidzana zvakanyanya, akaenzaniswa zvakasiyana, anoshandisa chikamu chikuru chawo patokeni yega yega yaanogadzira. Ndiyo sarudzo yeunyanzvi chaiyo, kwete kudzikisirwa.

Zvinotaurwa nenhamba dzeDeepSeek pachadzo

Makadhi epamutemo ese ari maviri anoburitsa zviyero pamiyero imwe cheteyo yemaagent neyeunjiniya hwesoftware, izvo zvinoita kuti uku kuve kuenzanisa kwakachena zvisingawanzoitiki — rabhoritari imwe chete, hurongwa hwekuongorora humwe chete, hukuru huviri:

Muyero (wakashumwa neDeepSeek)	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 inokunda pamutsara wega wega. Ndiwo musoro wechokwadi, uye tinofarira kuti uuverenge pano pane kuuwana muri kushandisa. Asi verenga mikaha, nokuti ndiyo yaunotenga chaizvo. PaTerminal Bench 2.1 mukaha wemapoinzi 5.2 pamodhi ine chikamu chechimakumi maviri neshanu chemapharamita; pakushandisa zvishandiso mapoinzi 3.8. Mukaha unowedzera apo basa richinyanya kudzama — mapoinzi 8.3 paDeepSWE, 7.3 paNL2Repo, ese ari maviri anopa mubairo kufungisisa kwekodhi kwematanho akawanda kunoenderera. Ndicho chimiro chekuchinjana mutafura imwe chete: Flash inovhara chikamu chikuru chenhambwe pamabasa akaganhurirwa uye inorasikirwa paasina muganhu.

Kadhi reFlash reDeepSeek pacharo rinojekesa zvakanyanya: Flash-0731 inokunda kuratidzwa kweV4-Pro pamiyero yakanyorwa pasinei nemapharamita ayo anoshanda ari mashoma zvikuru. Modhi itsva diki ichikunda huru yekare ndicho chinhu chakajairika mumunda uyu, uye ndicho chikonzero "mhando huru" isingafaniri kuva mhinduro yako yakagara iripo.

04Kuverenga chipfupiso: W4A16-AUTOROUND-REAP

Mazita edu emamodhi anotaura pachena hurongwa hwawo hwekudzvanya, nokuti ndicho chinhu chinonyanya kubata mhando yezvinobuda uye chinenge chisingataurwi nemunhu. Ichi chinokamurwa kuita nharo nhatu.

W4A16 — huremu hwe4-bit, maactivation e16-bit

Huremu hunochengetwa pamabhiti mana humwe nehumwe; maactivation anofamba nekururama kwe16-bit. Iyi i quantization yehuremu chete, uye inonangana nechinodzivirira chaicho pakugadzira matokeni, chinova bhandiwidhi yememori pane masvomhu. Kuita quantization pamaactivation zvakare kwaizowedzera kumhanya zvishoma nekuisa panjodzi kuwira kukuru kwekururama; kuasiya pa16 bits ndiko kusarudza kwekungwarira, kwekutendeseka kwepamusoro. Chinyorwa ichi chipfupiso chekushandisa — mabhuku e quantization anonyora chinhu chimwe chetecho seW4 nehukuru hweboka.

AUTOROUND — kutenderedza kwakadzidzwa, kwete kutenderedza kune iri pedyo

Quantization yakapusa inotenderedza huremu hwega hwega kukukosha kuri pedyo kunokwanisika. Izvozvo hazvina kunaka zvachose: kutenderedza huremu humwe nenzira "isiriyo" kunogona kudzima chikanganiso chakaunzwa nevavakidzani vahwo, uye pamabhiriyoni ehuremu kudyidzana ikoko kunotonga. AutoRound, kubva kuIntel, inotora sarudzo yekutenderedza sechinhu chekudzidza — kushanduka kunodzidziswa pahuremu humwe nehumwe kwakaganhurirwa pa±0.5, chaizvo zvakakwana kushandura sarudzo yekutenderedza ipi neipi imwe chete uye kwete kupfuura ipapo, pamwe nemiganhu yekucheka inodzidzwa, yakagadziriswa nesigned gradient descent ichirwisana nechinangwa chekuvakazve zvinobuda chikamu nechikamu paseti diki yekuyera. Matanho mazana maviri, wobva wapedza.

Kuenzanisa kwakabudiswa neIntel kune AutoRound ichikunda GPTQ mumamiriro 30 pa32 akaedzwa, AWQ mu27 pa32 neOmniQuant mu29 pa32 pazvirongwa zve4-bit zvehuremu chete, ine kunatsiridza pamusoro pekutenderedza kwakapusa kune iri pedyo kwemapoinzi ekururama 6.9 kusvika 33.2 pa2 bits. Hapana chimwe chazvo chinodhura panguva yekushandisa: kusarudza kwakanaka kwenhamba, kwakaitwa kamwe chete, kunze kwemutsetse.

REAP — kubvisa nyanzvi dzisingatumirwi nemunhu

REAP (Router-weighted Expert Activation Pruning), kubva kuCerebras Systems neUniversity of Calgary, ndiko kudzvanya kwakaitwa kamwe chete kwemamodhi eMoE. Inopa nyanzvi imwe neimwe chiyero maererano nekusimba nekuwanda kwekutumirwa kwayo nerouter, chichipetwa nehukuru hwezvinobuda zvenyanzvi iyoyo, ipapo yobvisa nyanzvi dzisingakoshi zvachose — pasina kudzidziswazve mushure.

Chinowanikwa chikuru chegwaro ndicho chinobatsira: pazviyero zvekudzvanya zvakakwira, kubvisa nyanzvi kunokunda kudzisanganisa. Kusanganisa nyanzvi dzinoita zvakasiyana kunokonzera zvinonzi nevanyori kuputsika kwenzvimbo yekushanda — router inorasikirwa nekudzora kwakatsetseka, kunotsamira pazvinopinda, kwakaita kuti MoE ikodzere kuvakwa, ichisara ichifambisa pakati pemaavheji akamira. Nhamba dzavo pamamodhi kubva pa20B kusvika 1T yemapharamita: pakudzvanywa kwenyanzvi kwe25% pamabasa ekukodha, REAP inorasikirwa ne2.8% yekururama kwepakati apo nzira dzekusanganisa dzichirasikirwa nekupfuura 5%; pa50%, REAP inorasikirwa ne8.0% apo kusanganisa kuchirasikirwa nekupfuura 20%. PaQwen3-Coder-480B pakubvisa kwe50% kudzikira kuri pa1.4%, kuenzaniswa ne39.3% yekubvisa kwakapusa kunotsamira pakuwanda. Cerebras yakaburitsa machekipoinzi eREAP emutsara weDeepSeek pakati pemamwe.

Sei 4-bit pano ne3-bit paPro

Mhando dzose dziri mbiri dzinofanira kukwana nekuramba dzichimhanya panhumbi dzerudzi rumwe chete. Pro inotakura mapharamita matiriyoni 1.7, saka inoda kubatwa kwakaomarara kwe3-bit kuti ikwane zvachose. Flash inotakura anenge mabhiriyoni mazana matatu, saka haizvidi — uye bhiti raanochengetedza rinoshandiswa pakutendeseka pane pakukwana.

Bhiti rekuwedzera iroro rinokosha kupfuura zvinoratidzwa nemasvomhu aro, nokuti kukuvara kwe quantization hakuna kutwasuka. Mukuyera kweWikiText-2 kwegwaro reAWQ, Llama-2-7B inobva pa5.47 perplexity paFP16 ichienda pa5.60 pa4-bit — chirango che2.4% — asi ichienda pa6.24 pa3-bit, chirango che14%. Ndiko kudzikira kunopetwa kanenge katanhatu nekuda kwekudzvanya kunowedzera ne25%, uye chimiro chimwe chetecho chinoenderera pa70B. Gwaro reGPTQ rakajeka zvakanyanya: pamabhiti matatu, kutenderedza kwakapusa kune iri pedyo kunoputsika pamamodhi erudzi rwe175B, kuchienda kuzviuru zveperplexity, asi kuchiramba kuchishanda zvakanaka pa4 bits. Pasi pemabhiti mana, algorithm yakanaka ye quantization inorega kuva chinhu chakanaka ichitanga kuva chinhu chega chakabata modhi pamwe chete.

Saka kuverenga kwechokwadi kwezvipfupiso izvi zviviri ndeuku: Flash iri pedyo neyayo isina kudzvanywa kupfuura Pro iri kune yayo. Pro ichiri modhi yakasimba — tafura yemiyero iri pamusoro inogadzirisa izvozvo. Yakangorasikirwa nezvakawanda kusvika pano, uye unobhadhara kuti uwane izvozvo.

Zvakatorwa kubva: 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. Nhamba dzeperplexity dzinobva kuAWQ (arXiv 2306.00978) neGPTQ (arXiv 2210.17323).

05Flash kupesana naPro pasuwo

Chikamu 03 chakafukidza hwaro. Heino mamiriro emhando mbiri sezvigadzirwa paGedhi redu:

	DeepSeek-V4-Flash-0731	DeepSeek-V4-Pro-0813
Quantization	W4A16-AUTOROUND-REAP	3BIT-REAP
Kururama kwehuremu	4-bit, kutenderedza kwakadzidzwa	3-bit
Hwaro (hwose / hunoshanda)	~304B / 13B	~1.7T / 49B
Anopinda pamatokeni 1M	$0.50	$1.95
Anobuda pamatokeni 1M	$2.00	$3.90
Context / zvinobuda zvakanyanya	262,144 / 16,384	262,144 / 16,384
Zvishandiso · zvinobuda zvakarongwa · kuyerera	Hongu	Hongu
Kuona	Kwete	Kwete
Chidzitiro chekuramba	Hapana	Hapana

Ona zvisina kuchinja: context, denga rezvinobuda, kudana zvishandiso, zvinobuda zvakarongwa, nekusavapo kwechidzitiro chekuramba zvakafanana. Zvinongochinja ihwaro, quantization, nemutengo. Izvozvo zvinoita kuti kuchinjana pakati padzo kuve shanduko yemutsara mumwe mukodhi yako uye sarudzo yemari chete muchivakwa chako.

Zvinodhurwa nekudana kumwe kwehukuru hwepamusoro

Zadza context yose yematokeni 262,144 utore zvinobuda zvakazara zvematokeni 16,384:

Pamutengo werondedzero. Flash: $0.131 anopinda + $0.033 anobuda. Pro: $0.511 + $0.064. Chiyero 3.51×. Flash inogovanawo mutengo wepasi we$0.50 / $2.00 neMiniMax-M3, Laguna-S-2.1, MiMo-V2.5, MiMo-V2.5-Pro neHy3 — ona katarogu yakazara kuti uone kusiyana kwadzo pakuona nepazvinobuda zvakarongwa.

06Masvomhu ehuwandu, akaitwa

"Yakachipa ka3.9" ichokwadi pane zvinopinda chete. Pane zvinobuda Flash yakachipa ka1.95. Kuchengetedza kwako chaiko naizvozvo kunotsamira pachiyero chako chezvinopinda kuzvinobuda, uye mukaha wakakura zvekuchinja sarudzo dzemagadzirirwo. Mabasa mana echokwadi, rimwe nerimwe pazvikumbiro 100,000:

Basa	Anopinda / anobuda pakudana	Flash	Pro	Kuchengetedza
Kurongedza matikiti	1,200 / 120	$84	$281	3.3×
Kupindura mibvunzo neRAG	8,000 / 600	$520	$1,794	3.4×
Pfupiso yegwaro refu	60,000 / 1,500	$3,300	$12,285	3.7×
Kugadzira dhata rekunyepedzera	500 / 2,000	$425	$878	2.1×

Chimiro: pauri kusundidzira context yakawanda uchigadzira mavara mashoma, ndipo Flash inonyanya kukuchengetedzera. Hurongwa hwekutsvaga, kurongedza magwaro, kuongorora zvinyorwa nemarogi, nekubvisa ruzivo pacontext refu ndiwo mabasa akanaka — anotongwa nezvinopinda kwazvo, uye zvinopinda ndipo pane kudzikiswa kwe3.9×. Kana pane izvozvo basa rako riri prompt pfupi inogadzira matokeni zviuru zvemavara, mukaha unodzikira kusvika kanenge kaviri uye nharo dzeunyanzvi dzePro dzinosimba zvakaenzana.

Chimwe chikamu chemasvomhu chinokodzera kunzwisiswa: pa$0.50 pamatokeni miriyoni anopinda, hwindo rose rematokeni 262,144 rinodhura masendi 13.1 kuzadza. Kuprompta necontext refu kunorega kuva mubvunzo webhajeti pamutengo iwoyo, izvo dzimwe nguva zvinoreva kuti unogona kusavaka danho rekutsvaga zvachose worega hwindo recontext richiita basa — kuchengetedza kweunjiniya pamusoro pekuchengetedza kwematokeni.

07Kuti "isina kudzorwa" inorevei pano, chaizvo

Shoko rinoshandiswa zvisina kunyatsotarwa, saka heino tsananguro yedu, yakataurwa sechinhu chesuwo: hapana muongorori wekuramba ari mberi kwemodhi uye hapana sefa yezvirimo inoshandiswa pane zvinobuda zvayo. Hapana chinobata prompt chodzosera kuramba kwakagadzirirwa; hapana chinoongorora mhinduro chichiivhara. Pabasa rehuwandu hukuru rekushandisa zvinobva ipapo hazvishamisi asi zvinobatsira kupfuura zvinoratidzwa neshoko racho:

• Hapana kuramba kusina musoro pahuwandu husina njodzi. Masuwo ane sefa anowanzoramba pfupiso dzezvokurapa, zvinyorwa zvezvekuchengetedza, kutsvaga humbowo hwepamutemo, ngano, zvinyorwa zvekudzikisa kukuvara nebasa rekudzora zvirimo pacharo — ndizvo chaizvo zvikamu apo basa rehuwandu riri kugadzirisa mavara ari pamusoro penyaya inonetsa pane kugadzira chinhu chinokuvadza. Pazvikumbiro zviuru zana, chiyero chekuramba kusina musoro cheimwe muzana mitsara chiuru yakaputsika.
• Zvibvumirano zvakatarwa. Kumbira mhinduro yejson_schema uwane JSON inoenderana neschema, kwete ndima yekukumbira ruregerero panofanira kuva nechinhu. Chiparadzanisi chako hachidi bazi rekuramba, uye pfungwa yako yekuedzazve haifaniri kusiyanisa kutadza kwetransport nekwemutemo.
• Ndiwe unotonga mutemo. Ndiko kuchinjana. Kana chigadzirwa chako chichida zvidziviriro zvevashandisi vako, unozvivaka padanho rako, apo unoziva vashandisi vako nedunhu rako, pane kugamuchira kufungidzira kwemupi pazviri zviviri. Akaundi yako inoramba ichitongwa neMutemo wedu Wekushandisa Nemutsigo.

Kusadzorwa hakusi kufanana nekusarongeswa kana kusabatsira. Modhi ichiri kutevedzera mirairo uye ichiri kuremekedza prompt yesisitimu. Kana uchida kuti irambe zvikamu zvakati kuvashandisi vako, zvitaure ipapo — murairo iwoyo zvino ndewako kunyora pane wedu kumanikidza.

08Unyanzvi nemiganhu yechokwadi

Unyanzvi huviri ndihwo hunotakura chikamu chikuru chebasa rehuwandu. Zvinobuda zvakasimba zvejson_schema zvinoshandura mhinduro kuita mutsara wedhatabhesi pasina danho rekuparadzanisa uye pasina bazi rekuramba. Kudana zvishandiso mumitauro mitatu yose kunoita Flash injini ine musoro yematanho madiki akachipa eagent ine muronga wepamusoro unomhanya pamodhi yakasimba. Wedzera reasoning_effort, inokutendera kushandisa matokeni ekuwedzera pachikamu chehuwandu chinoada chete, necontext yematokeni 262,144 yakafanana neyaPro — haurasikirwi necontext nekusarudza mhando yakachipa, izvo zvisingawanzoitiki.

Miganhu yakajeka zvakaenzana:

• Hapana kuona. Flash ndeyemavara chete uye icharamba zvikamu zvemifananidzo mumeseji. Kana uchida mifananidzo pamodhi isina kudzorwa pamutengo wepasi mumwe chete we$0.50 / $2.00, MiniMax-M3 neMiMo-V2.5 ndidzo sarudzo dzinoona; Kimi-K2.6 neKimi-K3 imamodhi anoona akasimba pamusoro pemhando iyoyo.
• Haisi pamusoro pemanera. Nhamba dzeDeepSeek pachadzo dzinoisa Pro mberi pamuyero wega wega wemaagent wakabudiswa, mukaha uri mukuru pakufungisisa kwekodhi kwematanho akawanda kunoenderera. Apo mhinduro imwe isiriyo inodhura kupfuura chiuru chakachipa, shandisa $1.45 yekuwedzera pamiriyoni paDeepSeek-V4-Pro.
• Matokeni 16,384 anobuda pachikumbiro. Zvinogadzirwa zvirefu zvinofanira kukamurwa pakudana kwakawanda. Iri ndiro denga rimwe chete pamodhi yega yega yatinoshandisa, uye rinosvikwa hwindo recontext risati rasvikwa.

09Kuidana

Nongedzera mutengi upi zvake anoenderana neOpenAI kuGedhi uchinje mutsara wemodhi. Hapana chimwe muchivakwa chako chinofanira kufamba.

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
        }
      }
    }
  }'

Zita remodhi rimwe chetero rinoshanda risina kuchinjwa pa/v1/messages ne/v1/responses, ndiyo nzira inokurumidza kupfuura dzose yekuedza modhi yakachipa pasi pekubatanidzwa kuripo usinganyorezve mutengi wako. Mapharamita, zvimiro zvezviitiko zvekuyerera, maschema ekudana zvishandiso nemiganhu yehuwandu zvakanyorwa murondedzero yeAPI; modhi yega yega mukatarogu yakanyorwa paShannon Research.

10Mibvunzo inowanzobvunzwa

---

Zvakatorwa kubva: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · Gwaro reunyanzvi reDeepSeek-V3 (arXiv 2412.19437) · DeepSeek-V4 (arXiv 2606.19348) · makadhi emamodhi eDeepSeek-V4-Flash-0731 neDeepSeek-V4-Pro-0813 · AutoRound (arXiv 2309.05516, Intel) · REAP (arXiv 2510.13999, Cerebras Systems) · AWQ (arXiv 2306.00978) · GPTQ (arXiv 2210.17323). Nhamba dzose dzehwaro nedzemiyero sezvadzakabudiswa nevanyori vadzo; mitengo yeShannon AI necontext yatinoshandisa ndedzedu dzerondedzero kusvika muna Gunyana 2026.