Ba ma ajiyewa ko samun damar bayanan kowane mai amfani, kuma ba ma dakatar da asusu sai idan hukuma ta doka ta buƙaci matakin aiwatarwa.
Samfura Marasa Tantancewa · Hy3 · W8A16

API na Samfurin Hy3 Marar Tantancewa

Tushen Hy3 mai 295B na Tencent a aminci na nauyi na 8-bit, tare da fitarwar json_schema mai tsauri, taga na tokens 262,144 kuma babu shimfidar ƙi — a farashi mafi ƙasa na katalogin.

An sabunta 5 ga Satumba, 2026Jagorar SamfuriTushe: Tencent Hy3Rubutu kaɗai

A Taƙaice

Hy3-W8A16 shi ne na uku a cikin samfuranmu uku na W8A16 — nauyi na 8-bit, kunnuwa na 16-bit, matakin mafi girman aminci da muke hidimarsa. Rubutu kaɗai ne, yana goyon bayan duka json_schema da response_format, yana ɗauke da mahalli na tokens 262,144 da rufin fitarwa na tokens 16,384 na duka katalogin, kuma yana kashe $0.50 / $2.00 kowane tokens miliyan — ƙasan. Babu shimfidar ƙi, babu matatar fitarwa. Ka zaɓe shi ga hanyoyin aikin rubutu masu yawa da dole ne su dawo da fitarwa mai nau'i, mai yiwuwar tantancewa a kowane kira, haɗe da a kan kayan da samfuri mai matata zai ƙi taɓawa.

Yawancin jayayyar "wane samfuri zan yi amfani da shi" tana game da rufi ne: wanne ne yake rubuta lamba mafi kyau, yake warware lissafi mafi wuya, yake nasara a gudun wakili mafi tsawo. Hanyoyin aiki masu tsari ba sa aiki haka. Aikin da yake cire filaye daga takardu dubu ɗari a rana ba ya buƙatar samfurin da yake da hazaka sau ɗaya — yana buƙatar wanda yake daidai a siffa guda a kowane lokaci, kuma wanda ba ya taɓa dawo da sakin layi inda abin JSON ya kamata ya kasance. Wannan ita ce siffar da Hy3-W8A16 yake nan don ita.

01Menene Hy3, kuma daga ina ya fito

Hy3 samfurin tunani da na wakilai ne mai buɗaɗɗen nauyi daga ƙungiyar Hunyuan ta Tencent, an sake shi gaba ɗaya a 6 ga Yuli, 2026 ƙarƙashin lasisi mai saukin kai, tare da nauyi da aka wallafa a Hugging Face. Yana ci gaba da zuriyar aikin Mixture-of-Experts na Hunyuan-Large na Tencent na baya. A gine-gine MoE mai raguwa ne: jimlar sigogi biliyan 295, wanda kusan biliyan 21 ke kunnuwa a kowane token ta turawa ta top-8 a kan masana 192, a kan transformer mai shimfidu 80 mai kulawa ta ƙungiyar tambaya, tare da shimfidar hasashen tokens da yawa mai 3.8B daban don ɓoye-ɓoye na hasashe. Mahalli na asali 256K ne — daidai tokens 262,144 da muke buɗewa, ba yankan wani abu mafi girma ba ko faɗaɗa wani abu ƙarami ta rope.

295B
Jimlar sigogi
21B
Masu kunnuwa / token
192
Masana (top-8)
262K
Taga mahalli

Wannan rabon raguwa ita ce lambar da take da muhimmanci ga duk wanda yake biya a kowane token: Hy3 yana ɗauke da ilimin samfuri mai 295B alhali yana yin kusan lissafin na 21B a kowace wucewa gaba. Shi ne kuma dalilin da ya sa za mu iya hidimarsa a 8 bits maimakon mu matsa shi da ƙarfi — ginin 8-bit na samfuri mai wannan raguwar har yanzu yana shiga kuma yana gudana da sauri a kan gungun GPU namu.

Abin da Tencent ya bayar da rahoton yake da kyau a kai

Teburan aunan da Tencent ya wallafa sun sanya ƙarfin Hy3 kai tsaye a cikin bincike na wakilai, tsara kayan aiki, da dawo da bayanai a dogon mahalli maimakon a rubuta lamba tsantsa. Lambobin da suka bayar da rahoto:

DeepSearchQA91.0
BrowseComp84.2
MCP-Atlas (kayan aiki)79.1
SWE-bench Verified78.0
AA-LCR (dogon mahalli)73.4
DeepSWE28.0

Muna nuna sandunan masu rauni da gangan. Hy3 yana bin GLM-5.2 a rubuta lamba a kwatancin Tencent da kansa — 78.0 idan aka kwatanta da 84.2 a SWE-bench Verified, 28.0 idan aka kwatanta da 46.2 a DeepSWE. Idan aikinka injiniyan software mai fayiloli da dama mai zaman kansa ne, ka karanta sauran fihirisar binciken samfura tukuna. Idan karanta rubutu ne da fitar da tsari, bayanin yana kusa da cikakke.

Lambar da ta fi muhimmanci ga aikin hanyar aiki ita ce ta aminci: Tencent ya bayar da rahoton adadin rikicewar Hy3 a 5.4%, ya sauka daga 12.5% a tsara ta baya, tare da kurakuran hankali na gama-gari sun sauka daga 25.4% zuwa 12.7%. Ƙa'idar horonsu da suka bayyana — a amsa idan akwai shaida, a faɗa a fili idan babu — ita ce ainihin ɗabi'ar da kake so a samfurin cirar bayanai wanda in ba haka ba zai ƙirƙiro daraja mai gamsarwa ga filin da tushen bai taɓa ambata ba.

Duk lambobin a wannan sashen Tencent ne ya bayar da rahotonsu daga gwaje-gwajensu; tabbatarwa mai zaman kanta ta ɓangare na uku har yanzu tana jira a lokacin rubutawa. Ka ɗauke su a matsayin alkibla, ba abin da aka sasanta ba.

02Menene ma'anar W8A16, kuma me ya sa yake da muhimmanci ga fitarwa mai tsari

Kowane samfuri a ƙofar shiga yana ɗauke da ƙarin matsi. Ba ado ba ne — yana gaya maka nisan da nauyin da ake hidimarsa yake daga wurin ajiyar da dakin binciken ya saki.

MatakiAbin da yake yiAminci
3BIT-REAPNauyi na 3-bit tare da REAP, hanyar datsa masana da take gogewa masanan MoE da ba a fi turawa gare su ba kafin matsiAn fi matsawa
W4A16-AUTOROUND-REAPNauyi na 4-bit / kunnuwa na 16-bit ta AutoRound — zagayen da aka koya wanda yake dawo da daidaito a kan zagaye zuwa mafi kusa mai sauƙi — tare da datsawa ta REAPTsakiya
W8A16Nauyi na 8-bit / kunnuwa na 16-bit. Babu datsawa, babu gyaran zagaye da aka koya — kawai ana ɗaukar nauyin a ninki biyu na daidaitoMafi girma

Inda kaucewar matsi take cutarwa da gaske

REAP da AutoRound suna dawo da mafi yawan abin da matsi mai sauƙi yake rasawa — amma "mafi yawa" tana yin aiki a wannan jimlar, kuma saura ba ya sauka daidai. Matsi na ƙananan bits yana rage wutsiyoyi: hukunce-hukunce masu ƙarancin yiwuwa, masu babban takamaimai. A cikin rubutu kusan ba a ganin hakan; wata siffa daban har yanzu jimla ce mai kyau. A cikin fitarwa mai tsari shi ne duka wasan, saboda schema doguwar sarƙa ce ta hukunce-hukunce masu ƙarancin yiwuwa da dole ne kowanne ya kasance daidai:

  • Nahawu a matakin token. Ƙarshen ɗaure, waƙafi, alamar faɗi. Kowanne token guda ne, kuma token guda mara kyau yana ɓata duka amsar.
  • Ladabin enum. Schema ɗinka ya ce "status" ɗaya ne daga cikin "open" | "closed" | "pending"; samfurin da ya kauce yana dawo da "Open" ko "pending review" — kusa a ma'ana, kuma ba daidai ba a ko ɗaya.
  • Ladabin nau'i. Filin number yana dawowa a matsayin "1,240", boolean a matsayin "yes". Duka biyun suna tantancewa a matsayin kirtani kuma suna faɗi a schema.
  • Kalmomin wutsiya masu tsawo. Sunaye masu kansu da ba a saba ba, lambobin sassa, sunayen sinadarai, rubutu ban da na Latin, kwanakin daidai — daidai fagen ƙarancin yiwuwa da nauyi na ƙananan bits yake rikitawa.

Babu abin da yake sa samfuran 4-bit ko 3-bit su zama marasa kyau; ga aikin tattaunawa bambancin aminci sau da yawa bai cancanci bambancin farashi ba. Amma idan mai amfani da fitarwar mai tabbatarwa ne maimakon mutum, daidaito a wutsiyoyi yana juyawa kai tsaye zuwa adadin tantancewa — kuma adadin tantancewa zuwa kasafin sake gwadawa da jinkirin p99. W8A16 shi ne matakin da ka daina biyan wannan harajin.

03Fitarwa mai tsari: json_schema da response_format

Hy3 yana goyon bayan duka rabin filin fitarwa mai tsari — ya cancanci a faɗe shi a fili, saboda hakan ba na kowa ba ne a katalogin:

Samfuriresponse_formatjson_schema
Hy3 · W8A16EhEh
MiMo-V2.5-Pro · W8A16EhEh
inkling · W4A16-AUTOROUND-REAPEhA'a
Laguna-S-2.1 · W4A16-AUTOROUND-REAPA'aA'a

Bambanci tsakanin hanyoyin biyun ya fi muhimmanci fiye da yadda mutane suke tsammani. response_format: {"type": "json_object"} yana neman JSON mai inganci kuma babu ƙari — samfurin yana zaɓar siffar, don haka kiraye-kiraye biyu za su iya dawo da {"name": ...} da {"full_name": ...} kuma duka biyun daidai ne. json_schema yana bayar da siffar tun farko: filaye masu suna, nau'ika, enum, jerin da ake buƙata, nannaɗewa. Wannan yarjejeniya ce da za ka iya rubuta lamba a kanta.

Ƙaramin kira a kan yaren mai jituwa da OpenAI:

POST /v1/chat/completions

{
  "model": "Hy3-W8A16",
  "messages": [
    {"role": "system", "content": "Extract fields from the incident report. Use null for anything not stated."},
    {"role": "user", "content": "<report text>"}
  ],
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "incident",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "severity":  {"type": "string", "enum": ["low","medium","high","critical"]},
          "component": {"type": "string"},
          "occurred_at": {"type": ["string","null"], "description": "ISO-8601"},
          "affected_users": {"type": ["integer","null"]},
          "indicators": {
            "type": "array",
            "items": {"type": "object", "properties": {
              "kind":  {"type": "string", "enum": ["ip","domain","hash","cve"]},
              "value": {"type": "string"}
            }, "required": ["kind","value"]}
          }
        },
        "required": ["severity","component","occurred_at","affected_users","indicators"]
      }
    }
  }
}

Bayanai masu amfani

  • Ka sanya kowane filin ya zama wajibi kuma ka ba da damar null. Filaye na zaɓi suna gayyatar faɗuwa a shiru; filin wajibi mai ba da damar null yana tilasta "ba ya nan" a fili, wanda bayani ne da za ka iya aiki da shi.
  • Enum sun fi kirtani kyauta. Duk inda za ka iya rufe tarin darajoji, ka rufe shi — canjin da ya fi tasiri a adadin nasarar schema a duniyar gaskiya.
  • Ka zauna ƙasa da rufin fitarwa na tokens 16,384. Tsarin da ba shi da iyaka zai yanke a tsakiyar abu. Ka iyakance tsawon tsari, ko ka raba shigarwar.
  • Ka riƙe saƙon tsari. Schema tana ƙuntata siffa; umarnin har yanzu yana gudanar da ma'ana — abin da ya zama "critical," wace kwanan wata ce ta yi nasara idan biyu suka bayyana, ko za a daidaita ma'aunai.
  • Ka tabbatar ko ta yaya. Goyon bayan schema yana ɗaga adadin nasara; ba ya maye gurbin mai tabbatarwa da layin saƙonnin da suka gaza.

04Me ya sa "marar tantancewa" siffa ce ta aminci, ba kawai ta abun ciki ba

Hy3 a ƙofar shiga yana da babu shimfidar ƙi kuma babu matatar fitarwa. Tsarin da aka saba shi ne game da abun ciki: za ka iya sarrafa binciken tsaro, nazarin malware, rubutun cin zarafin rauni, kayan likita da na shari'a, kayan manya, takardun rage cutarwa ko rubutun masu amfani masu ƙiyayya ba tare da samfurin ya yanke shawara a madadinka cewa ya fi son kada ya yi ba.

Akwai dalili na biyu, mafi na inji, da ya sa yake da muhimmanci. Ƙi karya schema ce. Samfuri mai matata da ya ƙi ba ya dawo da abinka da alamar "unavailable" — yana dawo da rubutu. "I'm not able to help with that request." Wannan kirtanin ba abin incident ba ne; yana faɗi a mai tabbatarwarka kuma yana sauka a layin saƙonnin da suka gaza kusa da ainihin kurakuran tantancewa, inda yanzu dole ne ka raba ƙin manufa da ainihin kurakuran tsari.

Mafi muni, ƙi tana da alaƙa da abun ciki maimakon ta rarraba da bazuwar. Idan kashi 4% na tarinka rahotannin abin da ya faru na tsaro ne kuma samfurin ya ƙi mafi yawansu, ba ka rasa kashi 4% na yawan aiki daidai ba — ka rasa rukuni guda, kuma dashboard ɗin da aka gina a kan waɗannan bayanan a shiru yana ƙarancin bayar da rahoton daidai abin da yake nan don ya fito da shi. Ramukan da suke da siffar rukuni sun fi tsada fiye da hayaniya iri ɗaya saboda babu abin da yake ba da faɗakarwa a kansu. Ka cire hanyar ƙin kuma abin da ya rage — kurakuran tsari, yanke, rubutu mai shubuha — ba shi da alaƙa da batun, don haka za a iya aunawa.

Marar tantancewa ba yana nufin babu lissafi ba

Babu shimfidar ƙi yana nufin samfurin ba ya yin sharhi a kan shigarwarka. Ba ya nufin ba a lura da asusun ba. Manufarmu ta Amfani da Alhaki da dokar da ta shafe ka ce take gudanar da amfani, kuma cin zarafi yana kawo ƙarshen shiga. Cire matatar yana hana samfuri yin tsammani a kan niyyarka daga ɗan ɓangaren takarda; ba ya cire alhakinka ga abin da ka gina.

05Mahalli, rufin fitarwa, da yadda za a tsara kasafinsu

Kowane samfuri a ƙofar shiga yana raba iyakoki guda:

262,144
Tokens na mahalli
16,384
Matsakaicin tokens na fitarwa
3
Yaruka na API
Eh
Streaming

Mahalli kasafi ne da ake rabawa: umarni tare da fitarwa dole ne su shiga cikin 262,144 tare, don haka rufin shigarwa mai amfani a kira guda kusan tokens 245,000 ne, kusan kalmomin Ingilishi 180,000. Ga aikin tsari rufin fitarwa na tokens 16,384 shi ne iyaka mafi wuya kuma wanda yake ba wa mutane mamaki — schema da take fitar da abu guda a kowane bayani tana bugansa tun kafin taga ta cika. Tsari biyu suna aiki a kan hakan:

  • Shigarwa mai faɗi, fitarwa mai kunkuntar. Ka tura duka kwangila, takardar shaida ko tagar log cikin mahalli ka nemi abu guda mai taƙaitawa. Tokens na shigarwa suna kashe kashi ɗaya cikin huɗu na tokens na fitarwa, don haka farashin yana ba da lada ga karanta da yawa da rubuta kaɗan.
  • Ka raba a kan iyakokin bayanai, ba adadin bytes ba. Ka raba inda samfurin zai iya ganin ɗinkin. Bayanin da aka yanke a tsakiyar jimla yana samar da abin da ba daidai ba da tabbaci maimakon gazawa bayyananne.

Ƙarfin dawo da bayanai a dogon mahalli na Hy3 da aka bayar da rahotonsa — 73.4 a AA-LCR, da tsalle daga 42.9% zuwa 75.1% a tunawa da doguwar tattaunawa ta MRCR — ita ce ainihin iyawar da tsari na farko ya dogara a kai, kuma ƙwarewa ce daban da rubutu mai kyau.

06Farashi mafi ƙasa, da lissafin hanyar aiki ta gaskiya

Hy3 yana zaune a $0.50 kowane tokens miliyan ɗaya na shigarwa da $2.00 kowane tokens miliyan ɗaya na fitarwa — ƙasan katalogin, ana rabawa da MiMo-V2.5, MiMo-V2.5-Pro, DeepSeek-V4-Flash-0731 da MiniMax-M3. An yi aiki ta nauyin wakilci:

AikiShigarwa / kiraFitarwa / kiraKuɗi / kiraKowane kira miliyan 1
Cirar filaye1,200300$0.0012$1,200
Rarraba takardu3,000120$0.0018$1,800
Taƙaita dogon takarda zuwa schema40,000900$0.0218$21,800
Nazarin cikakken taga240,0002,000$0.1240$124,000

Lambar da take canja hukunci ba babban farashin ba ce amma farashin mai tasiri bayan sake gwadawa. Idan samfura biyu suna kashe daidai kuma ɗaya yana gazawa a tabbatarwa a kashi 6% na kiraye-kiraye inda ɗayan yake gazawa a 1.5%, samfurin mai yawan kaucewa yana gudanar da kusan kashi 4.5% ƙarin kiraye-kiraye — abin jurewa. Barnar tana a wutsiya: kiraye-kirayen da aka sake gwadawa suna ɗaukar tsawon lokaci ninki biyu, don haka jinkirin p99 kusan yana ninkuwa ga kaso da ake gani na zirga-zirga, kuma duk abin da ya gaza sau biyu ya zama aikin hannu. A kiraye-kiraye miliyan a rana wannan dubban bayanai ne da wani dole ya karanta.

Saboda haka hujjar matakin 8-bit a farashi mafi ƙasa: ba kana biyan ƙari don aminci a nan ba, don haka lissafin sake gwadawa kuɗi ne kyauta. Idan samfurin 3-bit ya fi rahusa da za a sami ciniki na gaskiya da za a auna — Hy3 yana kashe daidai da abu mafi rahusa a katalogin.

07Yaruka uku na API, samfuri guda

Ana kai ga Hy3 a kan duka yaruka uku na ƙofar shiga tare da streaming a kowanne. Samfuri guda, nauyi guda, iyakoki guda — ka zaɓi wanda ya dace da SDK da kake da shi.

Ƙarshen hanyaYareA yi amfani da shi lokacin da
/v1/chat/completionsOpenAIKana da SDK na OpenAI, LangChain ko LlamaIndex. Ka nuna URL na asali a nan, ka canja alamar samfurin.
/v1/messagesAnthropicKana a kan SDK na Anthropic kuma kana son toshen abun ciki ba tare da shimfidar fassara ba.
/v1/responsesResponsesKana son sabuwar siffar buƙata da fitarwa mai nau'i da yanayin da uwar garke ke riƙewa.

Hy3 yana kuma goyon bayan kiran kayan aiki da tunani, haɗe da reasoning_effort. Ga aikin cirar bayanai matakin da ya fi amfani yakan zama a rage tunani: a kan schema da aka bayyana da kyau a kan rubutu mai tsafta, tokens na tunani tokens ne na fitarwa da kake biyansu da jinkirin da kake jira, don ƙaramin riba. Ka kashe kasafin ƙoƙarin a kan shigarwa mai shubuha da gaske maimakon haka.

Cikakken bayanin buƙata da amsa, jerin sigogi, ma'anonin streaming da wurin gwaji mai mu'amala suna cikin takardun API.

08Zaɓe tsakanin samfuran W8A16 ukun

Samfura uku suna zaune a matakin mafi girman aminci, duka a farashi mafi ƙasa. Zaɓin ya fi zama game da hanyar sadarwa:

SamfuriGanijson_schemaFarashi shiga / fitaKa zaɓe shi don
Hy3 · W8A16A'aEh$0.50 / $2.00Hanyoyin aiki masu tsari na rubutu kaɗai, tsara kayan aiki, dawo da bayanai a dogon mahalli
MiMo-V2.5-Pro · W8A16A'aEh$0.50 / $2.00Aikin rubutu kaɗai inda kake son daidaitawar Pro na layin MiMo
MiMo-V2.5 · W8A16EhEh$0.50 / $2.00Duk abin da ke da hotuna — takardun da aka duba, hotunan allo, hotuna

Idan wani ɓangare na shigarwarka hoto ne an riga an yanke maka hukunci: Hy3 rubutu kaɗai ne — ka aika PDF da aka duba, hotunan allo da hotuna zuwa MiMo-V2.5. Inda shigarwar rubutu ce da gaske, bayanin Hy3 na amfani da kayan aiki na wakilai, tunawa a dogon mahalli da ƙarancin rikicewar da aka bayar da rahotonsa ya danganta kai tsaye da aikin hanyar aiki. Tunda ukun suna kashe daidai, ka gudanar da schema ɗinka a kan biyu daga cikinsu a kan ƴan ɗari na ainihin bayanai ka kwatanta adadin tantancewa; wannan yana sasanta shi fiye da kowane tebur, haɗe da wannan.

09Inda Hy3 yake cin matsayinsa

  • Cirar filaye mai yawa. Kwangiloli, takardun shaida, rahotannin abin da ya faru, rubutun magana, bayanan asibiti, bayanan kotu — duk inda abin da ake bayarwa abu ne mai nau'i a kowace takarda kuma yawa yake sa sake gwadawa ya zama mai tsada.
  • Rarrabewa da turawa a girma. Fitarwar enum da aka rufe, inda duka darajar samfurin ita ce rashin taɓa ƙirƙirar rukuni na goma sha tara.
  • Tarin tsaro da bayanan barazana. Rubutun malware, bayyana cin zarafin rauni da samfuran phishing — kayan da suke sa samfura masu matatar su ƙi, kuma inda ramin da yake da siffar rukuni yake ɓata manufar.
  • Wakilan kiran kayan aiki da ba dole su tsaya ba. Amincin hujja na 8-bit yana nufin JSON da aka ba wa ayyukanka yana tantancewa; kiran kayan aiki da ya lalace wakili ne da ya tsaya, ba amsa mai ɗan muni ba.
  • Nazarin dogayen takardu zuwa tsari. Duka takardar shaida ta shiga, abu guda mai taƙaitawa ya fita.
  • Tantance abun ciki da binciken manufa. Yiwa kayan da samfuri mai matata ya ƙi karantawa lakabi matsala ce ta dindindin ta amana da tsaro; samfuri mara matata mai schema mai tsauri yana warware ta kai tsaye.

Kuma inda ba ya cin matsayinsa: duk abin da ke da hotuna (ka yi amfani da MiMo-V2.5), da injiniyan software mai fayiloli da dama mai zaman kansa, inda teburan Tencent da kansu suka sanya Hy3 a bayan GLM-5.2 a SWE-bench Verified kuma a baya sosai a DeepSWE.

Ka duba majiyoyin

Kowane iƙirari na gine-gine da na ma'auni game da tushen Hy3 yana zuwa daga kayan sakin jama'a na Tencent ko rahotannin manema labarai a kansu; gaskiyar katalogin tana zuwa daga /v1/models namu kai tsaye. Ka tabbatar da duka biyun.

Tencent ne ya bayar da rahoton lambobin ma'auni, haɗe da lambobin masu gasa; tabbatarwa mai zaman kanta tana jira a lokacin rubutawa. An lissafa lambobin kuɗi daga farashinmu da aka wallafa na $0.50 / $2.00. Babu lambar a wannan shafin da ta zama ma'aunin da Shannon ya gudanar ko ƙiyasi.

10Tambayoyin da ake yawan yi

Menene API na samfurin Hy3 marar tantancewa?

Ƙarshen hanyar Ƙofar Shiga ta Shannon AI ce ga alamar samfuri Hy3-W8A16 — tushen Hy3 mai buɗaɗɗen nauyi na Tencent ana hidimarsa a kan gungun GPU namu a daidaiton nauyi na 8-bit, ba tare da shimfidar ƙi ba kuma ba tare da matatar fitarwa ba. Yana bayar da taga mahalli na tokens 262,144, tokens 16,384 na fitarwa a kowace buƙata, kiran kayan aiki, tunani da reasoning_effort, da cikakkiyar fitarwa mai tsari ta duka json_schema da response_format, a kan /v1/chat/completions, /v1/messages da /v1/responses tare da streaming.

Shin Hy3 yana goyon bayan JSON schema da fitarwa mai tsari?

Eh. Hy3-W8A16 yana goyon bayan duka response_format ga yanayin abin JSON da json_schema ga schema masu tsauri, masu nau'i — cikakken filin fitarwa mai tsari, ba wani ɓangare ba. Wannan ba gaskiya ba ne ga kowane samfuri a katalogin: Laguna-S-2.1 ba ya goyon bayan ko ɗaya, kuma inkling yana goyon bayan response_format amma ba json_schema ba. Idan hanyar aikinka ta dogara ga yarjejeniya mai nau'i, Hy3 yana ɗaya daga cikin samfuran da za su iya girmama ta.

Menene ma'anar W8A16, kuma shin ya fi 4-bit ko 3-bit?

W8A16 yana nufin nauyi na 8-bit da kunnuwa na 16-bit — matakin mafi girman aminci da muke hidimarsa. Idan aka kwatanta da W4A16-AUTOROUND-REAP (nauyi na 4-bit, zagayen da aka koya, datsa masana) da 3BIT-REAP (nauyi na 3-bit tare da datsa masana), shi ne ya fi ƙarancin canja nauyin da aka saki, don haka ɗabi'a tana zama kusa da wurin ajiyar asali. Matakan masu tsanani suna nan saboda manyan tushe in ba haka ba ba za su zama masu yiwuwar hidima a sauri ba.

Shin Hy3 yana goyon bayan hotuna ko gani?

A'a. Hy3-W8A16 rubutu kaɗai ne. Aika ɓangarorin hoto zuwa gare shi ba zai yi aiki ba. Idan kana buƙatar gani tare da fitarwa mai tsari a matakin 8-bit guda, ka yi amfani da MiMo-V2.5-W8A16, wanda yake goyon bayan shigar hoto tare da json_schema da response_format a farashin $0.50 / $2.00 guda.

Nawa ne kuɗin API na Hy3?

$0.50 kowane tokens miliyan ɗaya na shigarwa da $2.00 kowane tokens miliyan ɗaya na fitarwa — farashi mafi ƙasa a katalogin mu, ana rabawa da MiMo-V2.5, MiMo-V2.5-Pro, DeepSeek-V4-Flash-0731 da MiniMax-M3. Hanyar aikin cirar bayanai ta wakilci a tokens 1,200 na shigarwa da 300 na fitarwa a kowane kira tana kashe kusan $0.0012 a kowane kira, ko kusan $1,200 kowane kira miliyan.

Shin Hy3 zai ƙi buƙata?

Babu shimfidar ƙi kuma babu matatar fitarwa a gaban samfurin. Ga hanyar aiki mai tsari wannan siffa ce ta aminci daidai da ta abun ciki: ƙi karya schema ce, saboda rubutun da yake bayyana dalilin da ya sa samfurin ba zai amsa ba ba ya tantancewa a matsayin abin da mai tabbatarwarka yake tsammani. Cire hanyar ƙin yana cire duka rukunin gazawar tantancewa a kan tarin tsaro, likita, shari'a da abun cikin manya.

Ka gudanar da Hy3 a kan schema ɗinka

Aminci na 8-bit, json_schema mai tsauri, mahalli 262K, babu ƙi — a farashi mafi ƙasa.

Karanta Takardun API Dukkan Binciken Samfura

Alamar samfuri Hy3-W8A16 · rubutu kaɗai · $0.50 shiga / $2.00 fita kowane tokens 1M


Majiyoyi: Katunan samfuri na Tencent Hunyuan · Ma'ajiyar Hy3 · Hunyuan-Large: An Open-Source MoE Model (arXiv 2411.02265) · Rahoton VentureBeat a kan sakin Hy3. Tencent ne ya bayar da rahoton ma'aunan tushe kuma na wucin gadi ne har sai an tabbatar da su da kai. Shannon Lab LLC, New Mexico, Amurka, ce take gudanar da Shannon AI.

Dukkan hanyoyin bincike