API na DeepSeek Flash Marar Tantancewa

Yawancin ƙungiyoyin da suke auna API na LLM suna inganta abin da bai dace ba. Suna auna a kan umarni ashirin mafi wuya da za su iya tunani, su zaɓi wanda ya yi nasara, sannan su gano bayan makonni shida cewa kashi casa'in da biyar na zirga-zirgarsu gajere ne, mai tsari, aiki mai maimaituwa da samfurin da yake kashe kaso ɗaya cikin huɗu na kuɗin zai sarrafa shi daidai. DeepSeek-V4-Flash yana nan don wannan kashi casa'in da biyar. Wannan shafin yana game da sanin lokacin da kake cikinsa — da lokacin da ba ka ciki ba.

01Abin da kake samu da gaske

Siffa	DeepSeek-V4-Flash-0731 · W4A16-AUTOROUND-REAP
Alamar samfuri	DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP
Mahalli / matsakaicin fitarwa	tokens 262,144 / tokens 16,384
Farashi kowane 1M (shigarwa / fitarwa)	$0.50 / $2.00
Streaming · kiran kayan aiki · tunani	Eh — yaruka ukun
Fitarwa mai tsari	json_schema da response_format
Shigar hoto (gani)	A'a — rubutu kaɗai
Shimfidar ƙi / matatar fitarwa	Babu
Ana hidimarsa a kan	Gungun GPU na Shannon

Yaruka uku na API suna kai ga nauyi guda: /v1/chat/completions (siffar OpenAI), /v1/messages (siffar Anthropic), da /v1/responses. Streaming yana aiki a kan ukun. Cikakken bayanin sigogi yana cikin takardun API.

02Daga ina ya fito: zuriyar DeepSeek

DeepSeek-AI ya kashe tsararraki biyar na samfura a kan sha'awa guda: samun ɗabi'ar gaba daga samfurin da yake da rahusa a hidima, ba kawai mai rahusa a horo ba. Kusan kowane hukuncin gine-gine a wannan layin yana komawa ga hakan, kuma an wallafa masu muhimmanci. Uku suna bayyana abin da kake kira lokacin da ka buga wannan ƙarshen hanyar.

DeepSeekMoE — masana masu daki-daki tare da guda ɗaya da ake rabawa

Shimfidar Mixture-of-Experts tana maye gurbin babbar hanyar sadarwa ta feed-forward guda da ƙanana da yawa da router da yake zaɓar ƴan kaɗan a kowane token. Nau'in DeepSeek yana yanka masana da daki-daki fiye da tsarin masana takwas na yau da kullum, don haka FLOPs masu kunnuwa guda suna sayen haɗi mai takamaimai sosai, kuma yana keɓe masani da ake rabawa wanda kowane token yake wucewa ta gare shi, don haka ba a sake koyan ilimin gama-gari a kowane ƙwararre ba. DeepSeek-V3 ya fito da hakan a matsayin masana 256 da ake turawa gare su tare da masani 1 da ake rabawa, ana kunna 8 da ake turawa gare su a kowane token — jimlar sigogi biliyan 671, biliyan 37 masu kunnuwa — tare da daidaita nauyi ba tare da asarar taimako ba: maimakon kalmar hukunci da take rage inganci, son zuciya na kowane masani yana ɗaga ko rage makin turawa a kowane mataki yayin da wannan masanin yake gudana da ƙasa ko sama da nauyinsa.

MLA — kulawar da take shiga cikin ƙwaƙwalwa

Ɗayan rabin matsalar kuɗin hidima shi ne ma'ajiyar KV, wadda take girma madaidaiciya da mahalli kuma tana mamaye ƙwaƙwalwa a tsawon jeri masu tsawo. Multi-head Latent Attention, wanda aka gabatar a DeepSeek-V2, yana matsa maɓallai da darajoji tare zuwa vector ɓoyayye mai ƙarancin matsayi kuma yana ajiye wancan, yana sake gina K/V mai amfani nan take. Lambobin V2 da DeepSeek ya wallafa, idan aka kwatanta da DeepSeek 67B ɗinsu na baya mai yawa, sun bayar da rahoton raguwar 93.3% a ma'ajiyar KV da yawan samarwa mafi girma sau 5.76×. Wannan shi ne dalilin da ya sa taga na tokens 262K abu ne na yau da kullum a wannan dangin maimakon ƙari mai tsada.

Kulawa mai raguwa — da tushen farashin tokens masu rahusa

V3.2-Exp ya ƙara DeepSeek Sparse Attention: mai fihirisa mai sauƙi yana ba wa tokens na mahalli maki idan aka kwatanta da tambayar, kuma mai zaɓe yana riƙe waɗanda suka fi maki kaɗai. Abin sha'awa shi ne shaidar, ba hanyar ba. DeepSeek ya wallafa kwatanci gefe-da-gefe da V3.1-Terminus inda MMLU-Pro yake daidai a 85.0, AIME 2025 ya haura (88.4 → 89.3), SWE-bench Verified ya sauka kaɗan (68.4 → 67.8) — inganci a tsaye kusan — kuma ya rage farashin API ɗinsu fiye da 50% a ranar guda. Gine-gine, da aka bayar da rahotonsa da gaskiya, ya juya kai tsaye zuwa farashi. Flash shi ne wannan hujjar da aka kai mataki ɗaya gaba.

Tsara ta V4

V4, da aka nuna a Afrilu 2026 ƙarƙashin lasisin MIT, ya sake maye gurbin tarin kulawa da haɗin Compressed Sparse Attention da Heavily Compressed Attention, yana horo da mai inganta Muon a kan tokens fiye da tiriliyan 32, kuma yana riƙe manufar hasashen tokens da yawa ta V3 — wadda take kuma aiki a matsayin ɓoye-ɓoye na hasashe, kuma wadda DeepSeek ya auna a adadin karɓar token na biyu na 85–90% don saurin ɓoye-ɓoye na kusan 1.8× a V3. Taƙaitaccen bayanin V4 ya bayar da rahoton V4-Pro yana gudana a kashi 27% na FLOPs na tunanin token guda na V3.2 da kashi 10% na ma'ajiyar KV ɗinsa. Dangin yana buɗe matakan ƙoƙarin tunani guda uku, wanda shi ne abin da sigar reasoning_effort ɗinmu take danganta da shi.

Majiyoyi: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · Rahoton Fasaha na DeepSeek-V3 (arXiv 2412.19437) · Katin samfuri na DeepSeek-V3.2-Exp · DeepSeek-V4 (arXiv 2606.19348). Duk lambobin DeepSeek ne ya bayar da rahotonsu.

03Menene ma'anar "Flash" a gaskiya

Wannan shi ne inda yawancin kwatanci suke kuskure, don haka ya cancanci a faɗe shi a fili: Flash na DeepSeek ba Pro ba ne da aka matsa, aka datse ko aka tace. Samfurin Mixture-of-Experts ne da aka horar da shi daban, ƙarami, mai nasa horon farko. Takardar nunin V4 ta bayar da girmajen biyu:

Lambobi daga takardar nunin DeepSeek-V4. Wuraren ajiyar hukuma da muke hidimarsu — V4-Flash-0731 da V4-Pro-0813 — sun jera jimloli ɗan girma na 304B da 1.7T a katunan samfurinsu; DeepSeek bai sake bayyana sigogi masu kunnuwa ga waɗannan ginin ba.

Lambobi biyu suna yin aikin a nan. Jimlar sigogi tana sanya sawun ƙwaƙwalwa, wanda yake yanke shawarar na'urar da kwafin samfurin zai mamaye. Sigogi masu kunnuwa suna sanya lissafi a kowane token, wanda yake yanke shawarar yawan aiki. Flash ya fi ƙanƙanta kusan sau 5.6× a na farko kuma sau 3.8× a na biyu — kuma akwai wani abu mai wayo da ya cancanci a lura da shi a wannan giɓin: Flash yana kunna 13 daga cikin sigogi biliyan 284, kusan 4.6%, alhali Pro yana kunna 49 daga cikin biliyan 1,600, kusan 3.1%. Flash shi ne samfurin da ya fi ƙarancin raguwa. Ba kawai Pro ne da ƙarancin masana a kunne ba; ƙira ce mai yawa, mai daidaito daban wadda take kashe kaso mafi girma na kanta a kan kowane token da take samarwa. Wannan zaɓi ne na injiniya na gaskiya, ba raguwa ba.

Abin da lambobin DeepSeek suke faɗa

Katunan samfuri na hukuma duka biyun suna wallafa maki a kan ma'aunan wakilai da na injiniyan software guda, wanda ya sa wannan ya zama kwatanci mai tsafta na musamman — dakin bincike guda, harsashin auna guda, girmaje biyu:

Ma'auni (DeepSeek ya bayar da rahoto)	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 ya yi nasara a kowane layi. Wannan shi ne kanun labarin gaskiya, kuma mun fi so ka karanta shi a nan maimakon ka gano shi a cikin samarwa. Amma ka karanta giɓin, saboda su ne abin da kake saya da gaske. A Terminal Bench 2.1 giɓin maki 5.2 ne a kan samfuri mai kashi ɗaya cikin ashirin da biyar na sigogi; a amfani da kayan aiki maki 3.8 ne. Giɓin yana faɗaɗa inda aikin ya fi zurfi — maki 8.3 a DeepSWE, 7.3 a NL2Repo, duka biyun suna ba da lada ga tunanin lamba mai matakai da dama mai dorewa. Wannan ita ce siffar cinikin a tebur guda: Flash yana rufe mafi yawan tazarar a kan ayyuka masu iyaka kuma yana bayar da gada a kan waɗanda ba su da iyaka.

Katin Flash na DeepSeek da kansa ya fitar da batun a fili sosai: Flash-0731 ya zarce nunin V4-Pro a kan ma'aunan da aka jera duk da ƙarancin sigogi masu kunnuwa sosai. Sabon ƙaramin samfuri da ya doke tsohon babba shi ne yanayin al'ada a wannan fannin, kuma shi ne dalilin da ya sa "mataki mafi girma" bai kamata ya taɓa zama amsarka ta asali ba.

04Karanta ƙarin: W4A16-AUTOROUND-REAP

Alamun samfuranmu suna faɗin tsarin matsinsu a fili, saboda shi ne abin da ya fi shafar ingancin fitarwa kuma kusan babu wanda yake bayyana shi. Wannan yana rabuwa zuwa iƙirari uku.

W4A16 — nauyi na 4-bit, kunnuwa na 16-bit

Ana adana nauyi a 4 bits kowanne; kunnuwa suna gudana a daidaiton 16-bit. Wannan matsi ne na nauyi kaɗai, kuma yana nufin ainihin ƙuntatawa a samar da tokens, wanda shi ne faɗin hanyar ƙwaƙwalwa maimakon lissafi. Matsa kunnuwa ma zai sayi ɗan ƙarin sauri kuma zai haɗari da faɗuwar daidaito mafi girma sosai; barinsu a 16 bits shi ne zaɓi mai taka tsantsan, mai babban aminci. Rubutun gajarcen tsarin hidima ne — littattafan matsi suna rubuta abu guda a matsayin W4 da girman ƙungiya.

AUTOROUND — zagayen da aka koya, ba zagaye zuwa mafi kusa ba

Matsi mai sauƙi yana juya kowane nauyi zuwa darajar da za a iya wakilta mafi kusa. Hakan a bayyane ba shi ne mafi kyau ba: juya nauyi ɗaya ta hanyar "mara kyau" zai iya soke kuskuren da maƙwabtansa suka gabatar, kuma da nauyi biliyoyin waɗannan mu'amaloli suna mamayewa. AutoRound, daga Intel, yana ɗaukar hukuncin zagayen kansa a matsayin abin da za a koya — canjin da za a iya horarwa a kowane nauyi wanda aka iyakance shi zuwa ±0.5, daidai gwargwadon isa a juya kowane hukuncin zagaye guda kuma ba fiye da haka ba, tare da iyakokin yanke da za a iya koya, wanda aka inganta ta gangarawar gradient mai alama a kan manufar sake gina fitarwa ta toshe-toshe a kan ƙaramin tarin daidaitawa. Matakai ɗari biyu, sannan an gama.

Kwatancen da Intel ya wallafa suna nuna AutoRound ya doke GPTQ a 30 daga cikin yanayi 32 da aka gwada, AWQ a 27 daga 32 da OmniQuant a 29 daga 32 a saitunan nauyi-kaɗai na 4-bit, da riba a kan zagaye zuwa mafi kusa mai sauƙi na maki 6.9 zuwa 33.2 na daidaito a 2 bits. Babu wani daga cikinsa da yake kashe komai a lokacin tunani: zaɓi ne mafi kyau na lambobi, da aka yi sau ɗaya, a waje da hanya.

REAP — goge masanan da babu wanda yake turawa gare su

REAP (Router-weighted Expert Activation Pruning), daga Cerebras Systems da Jami'ar Calgary, matsi ne na harbi guda ga samfuran MoE. Yana ba kowane masani maki bisa ƙarfin da router yake kai gare shi da sau nawa, an ninka shi da girman fitarwar wannan masanin, sannan ya goge masanan da suka fi ƙaranci kai tsaye — babu sake horo bayan haka.

Babban abin da takardar ta gano shi ne mai amfani: a girman matsi mai yawa, datsa masana ya fi haɗa su. Haɗa masana masu bambancin aiki yana haifar da abin da marubutan suke kira faɗuwar sararin aiki — router yana rasa ikon daki-daki, mai dogaro ga shigarwa wanda ya sa MoE ya cancanci ginawa, kuma an bar shi yana tuƙi tsakanin matsakaita marasa motsi. Lambobinsu a kan samfura daga 20B zuwa sigogi 1T: a matsi na masana na 25% a ayyukan rubuta lamba, REAP yana rasa matsakaicin daidaito na 2.8% inda ma'aunan haɗawa suke rasa fiye da 5%; a 50%, REAP yana rasa 8.0% inda haɗawa take rasa fiye da 20%. A Qwen3-Coder-480B a datsawar 50% raguwar 1.4% ce, idan aka kwatanta da 39.3% na datsawa mai sauƙi bisa yawan amfani. Cerebras ya wallafa wuraren ajiyar REAP ga layin DeepSeek a tsakanin sauran.

Me ya sa 4-bit a nan da 3-bit a Pro

Matakan biyun dole ne su shiga kuma su ci gaba da zama masu sauri a kan irin na'ura guda. Pro yana ɗauke da sigogi tiriliyan 1.7, don haka yana buƙatar maganin 3-bit mai tsanani don ya shiga kwata-kwata. Flash yana ɗauke da kusan biliyan 300, don haka ba ya buƙata — kuma ana kashe bit ɗin da ya adana a kan aminci maimakon a kan shiga.

Wannan ƙarin bit ɗin yana da muhimmanci fiye da yadda lissafinsa yake nunawa, saboda barnar matsi ba madaidaiciya ba ce. A cikin aunan WikiText-2 na takardar AWQ, Llama-2-7B yana tafiya daga rikicewar 5.47 a FP16 zuwa 5.60 a 4-bit — hukuncin 2.4% — amma zuwa 6.24 a 3-bit, hukuncin 14%. Wannan kusan ninki shida na raguwa don ƙarin matsi na 25% ne, kuma tsari guda yana riƙewa a 70B. Takardar GPTQ ta fi bayyana: a 3 bits, zagaye zuwa mafi kusa mai sauƙi yana faɗuwa a kan samfuran rukunin 175B, zuwa dubban rikicewa, alhali yana ci gaba da zama mai amfani sosai a 4 bits. A ƙasan bits huɗu, kyakkyawan algorithm na matsi ya daina zama abin ƙawa ya zama abu ɗaya tak da yake riƙe samfurin a haɗe.

Don haka karatun gaskiya na waɗannan ƙari biyun shi ne: Flash yana zaune kusa da kansa da ba a matsa ba fiye da yadda Pro yake zaune kusa da nasa. Pro har yanzu shi ne samfuri mafi ƙarfi — teburin ma'auni na sama ya sasanta hakan. Kawai ya bayar da ƙari don ya isa nan, kuma kana biyan kuɗin gatan.

Majiyoyi: 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. Lambobin rikicewa daga AWQ (arXiv 2306.00978) da GPTQ (arXiv 2210.17323).

05Flash da Pro a ƙarshen hanyar

Sashe na 03 ya rufe tushen. Ga yadda matakan biyun suke a matsayin kayayyaki a ƙofar shigarmu:

	DeepSeek-V4-Flash-0731	DeepSeek-V4-Pro-0813
Matsi	W4A16-AUTOROUND-REAP	3BIT-REAP
Daidaiton nauyi	4-bit, zagayen da aka koya	3-bit
Tushe (jimla / masu kunnuwa)	~304B / 13B	~1.7T / 49B
Shigarwa kowane tokens 1M	$0.50	$1.95
Fitarwa kowane tokens 1M	$2.00	$3.90
Mahalli / matsakaicin fitarwa	262,144 / 16,384	262,144 / 16,384
Kayan aiki · fitarwa mai tsari · streaming	Eh	Eh
Gani	A'a	A'a
Shimfidar ƙi	Babu	Babu

Ka lura da abin da ba ya canjawa: mahalli, rufin fitarwa, kiran kayan aiki, fitarwa mai tsari, da rashin shimfidar ƙi duk iri ɗaya ne. Abin da kaɗai yake motsi shi ne tushe, matsi, da farashi. Wannan ya sa canjawa tsakaninsu ya zama canjin kirtani guda a lambarka da hukunci na tattalin arziki kawai a gine-ginenka.

Nawa ne kudin kira guda mafi girma

Ka cika duka mahalli na tokens 262,144 ka kuma ɗauki cikakkiyar fitarwa ta tokens 16,384:

A farashin jeri. Flash: $0.131 shigarwa + $0.033 fitarwa. Pro: $0.511 + $0.064. Rabo 3.51×. Flash yana kuma raba ƙasan $0.50 / $2.00 da MiniMax-M3, Laguna-S-2.1, MiMo-V2.5, MiMo-V2.5-Pro da Hy3 — ka duba cikakken katalogin don yadda waɗannan suka bambanta a gani da fitarwa mai tsari.

06Lissafin yawa, an yi shi

"Rahusa sau 3.9×" gaskiya ne kawai a shigarwa. A fitarwa Flash ya fi rahusa sau 1.95×. Ainihin abin da ka adana don haka ya dogara ga rabon shigarwa zuwa fitarwa ɗinka, kuma faɗin ya isa ya canja hukunce-hukuncen ƙira. Ayyuka huɗu na gaskiya, kowanne a buƙatu 100,000:

Aiki	Shiga / fita kowane kira	Flash	Pro	Abin da aka adana
Rarraba tikiti	1,200 / 120	$84	$281	3.3×
Amsa tambaya ta RAG	8,000 / 600	$520	$1,794	3.4×
Taƙaita dogon takarda	60,000 / 1,500	$3,300	$12,285	3.7×
Samar da bayanan roba	500 / 2,000	$425	$878	2.1×

Tsarin: yawan mahallin da ka tura da ƙarancin rubutun da ka samar, shi ne yawan abin da Flash yake adana maka. Hanyoyin dawo da bayanai, tantance takardu, nazarin log da rubutun magana, da cirar bayanai a dogon mahalli su ne wurin da ya fi dacewa — shigarwa ce ta mamaye su da matakin girma, kuma shigarwa ce inda ragin sau 3.9× yake zaune. Idan a maimakon haka aikinka gajeren umarni ne da yake samar da dubban tokens na rubutu, giɓin yana matsawa zuwa kusan 2× kuma hujjar iyawa ta Pro tana ƙarfafa daidai gwargwado.

Wani lissafi ɗaya kuma da ya cancanci a haddace: a $0.50 kowane tokens miliyan ɗaya na shigarwa, duka taga na tokens 262,144 yana kashe cent 13.1 don a cika shi. Umarni a dogon mahalli ya daina zama tambayar kasafin kuɗi a wannan farashin, wanda wani lokaci yake nufin za ka iya tsallake gina shimfidar dawo da bayanai kwata-kwata ka bar taga mahalli ta yi aikin — adanawa ta injiniya a kan adanawar tokens.

07Menene ma'anar "marar tantancewa" a nan, daidai

Ana amfani da kalmar a sassauce, don haka ga ma'anarmu, an faɗe ta a matsayin siffar ƙarshen hanyar: babu mai rarrabewa na ƙi a gaban samfurin kuma babu matatar abun ciki da aka yi wa fitarwarsa. Babu abin da yake tsayar da umarni ya mayar da ƙi da aka riga aka rubuta; babu abin da yake duba amsar ya goge ta. Ga aikin samarwa mai yawa sakamakon ba shi da ban tsoro kuma ya fi amfani fiye da yadda kalmar take nunawa:

• Babu kuskuren tabbatarwa a kan zirga-zirga marar illa. Ƙarshen hanya masu matatar suna ƙin taƙaitawar likita, rubutun tsaro, gano shaida na shari'a, almara, kayan rage cutarwa da aikin tantancewa da kansa — daidai rukunan da aikin tsari yake sarrafa rubutu game da batu mai laushi maimakon ya samar da wani abu mai cutarwa. A buƙatu dubu ɗari, adadin ƙi na ƙarya na kashi ɗaya cikin ɗari yana nufin layuka dubu da suka lalace.
• Yarjejeniya masu tabbaci. Ka nemi amsar json_schema kuma ka samu JSON da ya dace da schema, ba sakin layi na neman afuwa a inda abu ya kamata ya kasance ba. Mai tantancewarka ba ya buƙatar reshen ƙi, kuma dabarunka na sake gwadawa ba dole ne su bambanta gazawar tafiya da ta manufa ba.
• Kai ne ke da manufar. Wannan shi ne cinikin. Idan kayanka yana buƙatar shinge ga masu amfani da shi na ƙarshe, ka gina su a shimfidarka, inda ka san masu amfani da kai da hukuncinka, maimakon ka gaji tsammanin mai bayarwa game da su biyun. Asusunka ya kasance ƙarƙashin Manufarmu ta Amfani da Alhaki.

Marar tantancewa ba daidai yake da rashin daidaituwa ko rashin taimako ba. Samfurin har yanzu yana bin umarni kuma yana girmama umarnin tsari. Idan kana son ya ƙi wasu rukunai ga masu amfani da kai, ka faɗa a nan — wannan umarnin yanzu naka ne ka rubuta maimakon namu mu ɗora.

08Iyawa da iyakoki na gaskiya

Iyawa biyu suna ɗaukar mafi yawan nauyi ga aikin yawa. Fitarwar json_schema mai tsauri tana juya amsa zuwa layin bayanai ba tare da shimfidar tantancewa ba kuma ba tare da reshen ƙi ba. Kiran kayan aiki a yaruka ukun yana sa Flash injin da ya dace ga ƙananan matakai masu rahusa na wakilin da mai tsara shirinsa na sama yake gudana a kan samfuri mafi ƙarfi. Ka ƙara reasoning_effort, wanda yake ba ka damar kashe ƙarin tokens kawai a kan kaso na zirga-zirgar da yake buƙatarsu, da mahalli na tokens 262,144 daidai da na Pro — ba ka bar mahalli ta zaɓar matakin mai rahusa ba, wanda abu ne da ba a saba gani ba.

Iyakokin suna da tabbas daidai:

• Babu gani. Flash rubutu kaɗai ne kuma zai ƙi ɓangarorin hoto a cikin saƙo. Don hotuna a kan samfuri marar tantancewa a ƙasan $0.50 / $2.00 guda, MiniMax-M3 da MiMo-V2.5 su ne zaɓuɓɓukan masu iya gani; Kimi-K2.6 da Kimi-K3 su ne samfuran gani mafi ƙarfi a sama da wannan matakin.
• Ba shi ne saman tsanin ba. Lambobin DeepSeek da kansu sun sanya Pro a gaba a kowane ma'auni na wakilai da aka wallafa, inda giɓin ya fi faɗi a tunanin lamba mai matakai da dama mai dorewa. Inda amsa mara kyau guda take kashe fiye da dubu masu rahusa, ka kashe ƙarin $1.45 kowane miliyan a kan DeepSeek-V4-Pro.
• Tokens 16,384 na fitarwa a kowace buƙata. Dole ne a raba samarwa masu tsawo a kan kiraye-kiraye. Wannan rufi guda ne a kowane samfuri da muke hidimarsa, kuma yana ɗaure tun kafin taga mahalli ta yi.

09Kiransa

Ka nuna duk wani abokin ciniki mai jituwa da OpenAI zuwa ƙofar shigar ka canja kirtanin samfurin. Babu wani abu a tarinka da dole ya motsa.

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
        }
      }
    }
  }'

Alamar samfuri guda tana aiki ba canjawa a kan /v1/messages da /v1/responses, wadda ita ce hanya mafi sauri ta gwada samfuri mai rahusa a ƙarƙashin haɗin da ake da shi ba tare da sake rubuta abokin cinikinka ba. An rubuta sigogi, siffofin abubuwan streaming, schema na kiran kayan aiki da iyakokin adadi a cikin bayanin API; an jera kowane samfuri a katalogin a Shannon Research.

10Tambayoyin da ake yawan yi

---

Majiyoyi: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · Rahoton Fasaha na DeepSeek-V3 (arXiv 2412.19437) · DeepSeek-V4 (arXiv 2606.19348) · Katunan samfuri na DeepSeek-V4-Flash-0731 da DeepSeek-V4-Pro-0813 · AutoRound (arXiv 2309.05516, Intel) · REAP (arXiv 2510.13999, Cerebras Systems) · AWQ (arXiv 2306.00978) · GPTQ (arXiv 2210.17323). Duk lambobin tushe da na ma'auni kamar yadda marubutansu suka wallafa su; farashin Shannon AI da mahallin da muke hidimarsa darajojinmu ne na jeri tun Satumba 2026.