We do not hold or access any user's data, and we do not suspend accounts unless a lawful authority requires an enforcement action.
Mitundu Yosasidwa · Hy3 · W8A16

API Yosasidwa ya Mtundu wa Hy3

Maziko a Hy3 a 295B a Tencent pa kukhulupirika kwa zolemera kwa 8-bit, ndi zotuluka zolimba za json_schema, zenera la ma token 262,144 ndi palibe wosanjikiza wa kukana — pa mtengo wa pansi wa katalogi.

Yasinthidwa Seputembala 5, 2026Chitsogozo cha MtunduMaziko: Tencent Hy3Mawu okha

Kufupikitsa

Hy3-W8A16 ndi wachitatu mwa mitundu yathu itatu ya W8A16 — zolemera za 8-bit, zochitika za 16-bit, mlingo wa kukhulupirika kwapamwamba kwambiri umene timatumikira. Ndi wa mawu okha, umathandiza zonse json_schema ndi response_format, umanyamula context ya ma token 262,144 ya katalogi yonse ndi malire a zotuluka a ma token 16,384, ndipo umalipiritsa $0.50 / $2.00 pa miliyoni ya ma token — pansi penipeni. Palibe wosanjikiza wa kukana, palibe kusefa kwa zotulutsa. Musankheni pa njira za mawu za kuchuluka kwakukulu zomwe ziyenera kubweza zotuluka zowerengeka, zokhala ndi mtundu wodziwika pa yitanidwe lililonse, kuphatikiza pa zinthu zimene mtundu wosefedwa ukanakana kuzikhudza.

Mikangano yambiri ya "ndi mtundu uti umene ndiyenera kugwiritsa ntchito" ndi ya malire apamwamba: ndi uti umene umalemba code yabwino kwambiri, umathetsa masamu ovuta kwambiri, umapambana kuyenda kwautali kwa agent. Njira zokonzedwa za deta sizigwira ntchito motero. Ntchito yotulutsa magawo mu zikalata zana masauzande pa tsiku siifuna mtundu umene ndi wanzeru kamodzi — imafuna umene uli wolondola mu mawonekedwe omwewo nthawi zonse, ndipo umene sudzabweza ndime pomwe payenera kukhala chinthu cha JSON. Ndicho khalidwe limene Hy3-W8A16 ilipo pano.

01Hy3 ndi chiyani, ndipo imachokera kuti

Hy3 ndi mtundu wa zolemera zotseguka wa kuganiza ndi agentic wochokera ku gulu la Hunyuan la Tencent, wotulutsidwa mokwanira pa July 6, 2026 pansi pa layisensi yololera, ndi zolemera zofalitsidwa pa Hugging Face. Umapitiriza mzere wa ntchito yakale ya Hunyuan-Large ya Mixture-of-Experts ya Tencent. Mwa kamangidwe ndi MoE yosakanikirana: zigawo zonse mabiliyoni 295, mwa izo pafupifupi mabiliyoni 21 amayatsidwa pa token iliyonse kudzera mu routing ya top-8 pa akatswiri 192, pa transformer ya zosanjikiza 80 ndi grouped-query attention, kuphatikiza chosanjikiza chapadera cha multi-token-prediction cha 3.8B cha speculative decoding. Context yeniyeni ndi 256K — ndendende ma token 262,144 amene timapereka, osati kudulidwa kwa chinthu chachikulu kapena kutambasulidwa kwa rope kwa chinthu chaching'ono.

295B
Zigawo zonse
21B
Zogwira ntchito / token
192
Akatswiri (top-8)
262K
Zenera la context

Chiŵerengero cha kusakanikirana ndicho nambala yofunika kwa aliyense amene akulipira pa token: Hy3 imanyamula chidziwitso cha mtundu wa 295B pomwe ikuchita pafupifupi masamu a wa 21B pa kudutsa kulikonse. Ndicho chifukwanso chake tingathe kuitumikira pa mabiti 8 m'malo moipanikiza mwamphamvu — kupanga kwa 8-bit kwa mtundu wosakanikirana chonchi kukadali kokwana ndipo kukadali kothamanga pa gulu lathu la ma GPU.

Chomwe Tencent ikunena kuti ndi yabwino pa chiyani

Matebulo a kuyeza ofalitsidwa a Tencent akuika mphamvu za Hy3 mwachindunji pa kufufuza kwa agentic, kulinganiza zida, ndi kubwezeretsa pa context yaitali m'malo mwa kulemba code kokha. Ziwerengero zawo zofalitsidwa:

DeepSearchQA91.0
BrowseComp84.2
MCP-Atlas (zida)79.1
SWE-bench Verified78.0
AA-LCR (ctx yaitali)73.4
DeepSWE28.0

Timasonyeza mizere yofooka mwadala. Hy3 imatsalira GLM-5.2 pa kulemba code mu kuyerekezera kwa Tencent kokha — 78.0 motsutsana ndi 84.2 pa SWE-bench Verified, 28.0 motsutsana ndi 46.2 pa DeepSWE. Ngati ntchito yanu ndi uinjiniya wa mapulogalamu wa mafayilo ambiri wodziyendetsa, werengani zotsala za mndandanda wa kafukufuku wa mitundu choyamba. Ngati ndi kuwerenga mawu ndi kutulutsa mawonekedwe, mbiri yake ili pafupi ndi yangwiro.

Chiwerengero chofunika kwambiri pa ntchito ya njira za deta ndi cha kudalirika: Tencent ikunena za mlingo wa kupeka wa Hy3 pa 5.4%, kutsika kuchokera pa 12.5% mu mbadwo wakale, ndi zolakwika za nzeru zamba zikutsika kuchokera pa 25.4% mpaka 12.7%. Mfundo yawo yonenedwa ya kuphunzitsa — yankhani pamene pali umboni, nenani momveka pamene palibe — ndendende ndi khalidwe limene mukufuna mu mtundu wotulutsa deta umene ukanapanga mtengo wooneka ngati woona wa gawo limene gwero silinalitchule konse.

Ziwerengero zonse mu gawoli ndi zonenedwa ndi Tencent kuchokera ku mayeso awo okha; kutsimikizira kodziyimira pawokha kunali kukudikirabe pa nthawi yolembera. Zitengeni ngati zosonyeza njira, osati zothetsedwa.

02Chomwe W8A16 chikutanthauza, ndi chifukwa chake chikufunika pa zotuluka zokonzedwa

Mtundu uliwonse pa gateway umanyamula chowonjezera cha quantization. Si chokongoletsa — chimakuuzani mtunda wa zolemera zotumikiridwa kuchokera ku checkpoint imene labotale inatulutsa.

MlingoChomwe umachitaKukhulupirika
3BIT-REAPZolemera za 3-bit kuphatikiza REAP, njira yodula akatswiri imene imachotsa akatswiri a MoE omwe amagawiridwa ntchito kawirikawiri isanachite quantizationWopanikizidwa kwambiri
W4A16-AUTOROUND-REAPZolemera za 4-bit / zochitika za 16-bit kudzera mu AutoRound — kuzungulira kophunziridwa kumene kumabweza kulondola pamwamba pa kuzungulira ku pafupi mosaganizira — kuphatikiza kudula kwa REAPPakati
W8A16Zolemera za 8-bit / zochitika za 16-bit. Palibe kudula, palibe kukonza kwa kuzungulira kophunziridwa — zolemera zimangonyamulidwa pa kulondola kowirikiza kawiriKwapamwamba kwambiri

Kumene kusochera kwa quantization kumapweteka kwenikweni

REAP ndi AutoRound zimabweza zambiri mwa zomwe kupanikiza mosaganizira kumataya — koma "zambiri" likugwira ntchito mu chiganizo chimenecho, ndipo zotsala sizigwera mofanana. Quantization ya mabiti otsika imawononga michira: zisankho za kuthekera kotsika, za kudziwika kwakukulu. Mu ndime izi zimakhala pafupifupi zosaoneka; liwu lofotokoza losiyana likadali chiganizo chabwino. Mu zotuluka zokonzedwa ndicho chinthu chonse, chifukwa schema ndi unyolo wautali wa zisankho za kuthekera kotsika zimene zonse ziyenera kukhala zolondola ndendende:

  • Kalembedwe ka mlingo wa token. Chotsekera, koma, chizindikiro cha mawu. Chilichonse ndi token imodzi, ndipo token imodzi yolakwika imawononga yankho lonse.
  • Kutsatira kwa enum. Schema yanu ikunena kuti "status" ndi imodzi mwa "open" | "closed" | "pending"; mtundu wosochera umabweza "Open" kapena "pending review" — zoyandikira pa tanthauzo, ndipo sizovomerezeka pa zonse ziwiri.
  • Kutsatira kwa mtundu wa deta. Gawo la number limabwera ngati "1,240", la boolean ngati "yes". Zonse ziwiri zimawerengedwa ngati mawu ndipo zimalephera schema.
  • Mawu a mchira wautali. Mayina enieni osowa, manambala a zigawo, mayina a mankhwala, zilembo zosakhala za Chilatini, masiku olondola — ndendende dera la kuthekera kotsika limene zolemera za mabiti otsika zimasokoneza.

Palibe chilichonse mwa izi chimene chimapangitsa mitundu ya 4-bit kapena 3-bit kukhala yoipa; pa ntchito ya kukambirana kusiyana kwa kukhulupirika nthawi zambiri sikuyenera kusiyana kwa mtengo. Koma pamene wolandira zotulukazo ndi chotsimikizira osati munthu, kulondola pa michira kumasandulika mwachindunji kukhala mlingo wa kuwerengedwa — ndipo mlingo wa kuwerengedwa kukhala ndalama za kuyesanso ndi p99 latency. W8A16 ndi mlingo umene mumasiya kulipira msonkho umenewo.

03Zotuluka zokonzedwa: json_schema ndi response_format

Hy3 imathandiza zonse ziwiri za zotuluka zokonzedwa — kufunika kunena momveka, chifukwa sizili choncho pa katalogi yonse:

Mtunduresponse_formatjson_schema
Hy3 · W8A16IndeInde
MiMo-V2.5-Pro · W8A16IndeInde
inkling · W4A16-AUTOROUND-REAPIndeAyi
Laguna-S-2.1 · W4A16-AUTOROUND-REAPAyiAyi

Kusiyana pakati pa njira ziwirizi kumafunika kuposa mmene anthu amaganizira. response_format: {"type": "json_object"} imapempha JSON yovomerezeka basi — mtunduwo umasankha mawonekedwe, choncho mayitanidwe awiri angabweze {"name": ...} ndi {"full_name": ...} ndipo onse awiri ndi olondola. json_schema imapereka mawonekedwe patsogolo: magawo otchulidwa, mitundu ya deta, ma enum, mindandanda yofunikira, kulukidwa. Icho ndi pangano limene mungalembere code.

Yitanidwe laling'ono pa chinenero chogwirizana ndi 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"]
      }
    }
  }
}

Zolemba zothandiza

  • Pangani gawo lililonse kukhala lofunikira ndipo lololeze null. Magawo osankha amayitanitsa kuchotsedwa mwakachetechete; gawo lofunikira lololeza null limakakamiza kunena momveka kuti "kulibe," chimene ndi chidziwitso chimene mungachichitire kanthu.
  • Ma enum amapambana mawu aulere. Kulikonse kumene mungatseke gulu la mitengo, litseke — kusintha kwa mphamvu kwambiri pa mlingo weniweni wa kudutsa wa schema.
  • Khalani pansi pa malire a ma token 16,384 otuluka. Mndandanda wopanda malire udzadulika pakati pa chinthu. Ikani malire a kutalika kwa mindandanda, kapena gawani chilowetso.
  • Sungani uthenga wa dongosolo. Schema imalamulira mawonekedwe; funso likadali kulamulira tanthauzo — chomwe chimatchedwa "critical," tsiku liti limapambana pamene awiri akuoneka, ngati muyenera kusintha miyeso.
  • Tsimikizani mulimonse. Kuthandiza kwa schema kumakweza mlingo wa kudutsa; sikulowa m'malo mwa chotsimikizira ndi mzere wa makalata akufa.

04Chifukwa chake "chosasidwa" ndi khalidwe la kudalirika, osati la zinthu zokha

Hy3 pa gateway ilibe wosanjikiza wa kukana ndipo ilibe kusefa kwa zotulutsa. Kanenedwe kake ka nthawi zonse ndi ka zinthu: mungakonze kafukufuku wa chitetezo, kusanthula malware, zolemba za kuzembera, zolemba za zachipatala ndi zamalamulo, zinthu za akuluakulu, zolemba za kuchepetsa kuwonongeka kapena zolembedwa za ogwiritsa ntchito ankhanza popanda mtunduwo kusankha m'malo mwanu kuti ukanakonda kusatero.

Pali chifukwa chachiwiri, cha makina kwambiri, chimene chimapangitsa izi kufunika. Kukana ndi kuphwanya kwa schema. Mtundu wosefedwa umene umakana subweza chinthu chanu ndi chizindikiro cha "unavailable" — umabweza ndime. "I'm not able to help with that request." Mzere umenewo si chinthu cha incident; umalephera chotsimikizira chanu ndipo umagwera mu mzere wa makalata akufa pambali pa zolakwika zenizeni za kuwerenga, kumene tsopano muyenera kulekanitsa kukana kwa ndondomeko ndi zolakwika zenizeni za mawonekedwe.

Choipa kwambiri, kukana kumagwirizana ndi zinthu m'malo mogawidwa mwangozi. Ngati 4% ya zolemba zanu ndi malipoti a zochitika za chitetezo ndipo mtunduwo umakana ambiri a iwo, simunataye 4% ya kuchuluka mofanana — mwataya gulu limodzi, ndipo dashboard yomangidwa pa deta imeneyo mwakachetechete imanena mochepa za chinthu chomwecho chimene ilipo kuti ichiwonetse. Mabowo a mawonekedwe a magulu amawononga kuposa phokoso lofanana chifukwa palibe chimene chimachenjeza pa iwo. Chotsani njira ya kukana ndipo zotsala — zolakwika za mawonekedwe, kudulika, mawu osamveka bwino — sizigwirizana ndi mutu, ndipo choncho zingayezedwe.

Chosasidwa sichitanthauza chopanda udindo

Kusakhala kwa wosanjikiza wa kukana kukutanthauza kuti mtunduwo suweruza za chilowetso chanu. Sikutanthauza kuti akaunti siiyang'aniridwa. Kagwiritsidwe kayendetsedwa ndi Ndondomeko yathu ya Kagwiritsidwe Kaudindo ndi lamulo logwira ntchito, ndipo kuzunza kumathetsa mwayi. Kuchotsa chosefera kumaletsa mtundu kuyerekezera cholinga chanu kuchokera ku kagawo ka chikalata; sikuchotsa udindo wanu pa zomwe mukumanga.

05Context, malire a zotuluka, ndi mmene mungazikonzere

Mtundu uliwonse pa gateway umagawana malire omwewo:

262,144
Ma token a context
16,384
Ma token otuluka kwambiri
3
Zilankhulo za API
Inde
Streaming

Context ndi ndalama zogawana: funso kuphatikiza zotuluka ziyenera kukwana mkati mwa 262,144 pamodzi, choncho malire enieni a chilowetso mu yitanidwe limodzi ali pafupifupi ma token 245,000, mopanda kulondola kwambiri mawu 180,000 a Chingerezi. Pa ntchito yokonzedwa malire a ma token 16,384 otuluka ndiwo malire ovuta kwambiri ndipo ndiwo amene amadabwitsa anthu — schema yotulutsa chinthu chimodzi pa mbiri iliyonse imawakhudza zenera lisanadzaze. Mawonekedwe awiri amagwira ntchito motsutsana ndi izo:

  • Chilowetso chachikulu, zotuluka zochepa. Kankhirani pangano lonse, chikalata kapena zenera la ma log mu context ndipo pemphani chinthu chimodzi chachifupi. Ma token olowa amalipiritsa gawo limodzi mwa magawo anayi a ma token otuluka, choncho mtengo umapatsa mphotho kuwerenga zambiri ndi kulemba pang'ono.
  • Gawani pa malire a mbiri, osati pa kuchuluka kwa ma byte. Gawani kumene mtunduwo ungathe kuona mgwirizano. Mbiri yodulidwa pakati pa chiganizo imapanga chinthu cholakwika modzidalira m'malo mwa kulephera koonekera.

Mphamvu yonenedwa ya Hy3 ya kubwezeretsa pa context yaitali — 73.4 pa AA-LCR, ndi kudumpha kuchokera 42.9% mpaka 75.1% pa kukumbukira kwa kukambirana kwautali kwa MRCR — ndendende ndi luso limene mawonekedwe oyamba amadalira, ndipo ndi luso losiyana ndi kulemba bwino.

06Mtengo wa pansi, ndi masamu a njira yeniyeni ya deta

Hy3 ili pa $0.50 pa miliyoni ya ma token olowa ndi $2.00 pa miliyoni ya ma token otuluka — pansi pa katalogi, pogawana ndi MiMo-V2.5, MiMo-V2.5-Pro, DeepSeek-V4-Flash-0731 ndi MiniMax-M3. Zowerengedwa pa ntchito zoyimira:

NtchitoKulowa / yitanidweKutuluka / yitanidweMtengo / yitanidwePa mayitanidwe 1M
Kutulutsa magawo1,200300$0.0012$1,200
Kuika zikalata m'magulu3,000120$0.0018$1,800
Chidule cha chikalata chachitali kupita ku schema40,000900$0.0218$21,800
Kusanthula kwa zenera lonse240,0002,000$0.1240$124,000

Nambala imene imasintha zisankho si mtengo wa mutu koma mtengo weniweni pambuyo pa kuyesanso. Ngati mitundu iwiri ikulipiritsa chimodzimodzi ndipo umodzi umalephera kutsimikizira pa 6% ya mayitanidwe pomwe wina umalephera pa 1.5%, mtundu wosochera kwambiri umayendetsa mayitanidwe pafupifupi 4.5% owonjezera — kolekereka. Kuwonongeka kuli pa mchira: mayitanidwe obwerezedwa amatenga nthawi yowirikiza kawiri, choncho p99 latency imawirikiza kawiri pa gawo looneka la magalimoto, ndipo chilichonse chimene chimalephera kawiri chimakhala ntchito ya manja. Pa mayitanidwe miliyoni pa tsiku izo ndi mbiri masauzande zimene wina ayenera kuwerenga.

Choncho nkhani ya mlingo wa 8-bit pa mtengo wa pansi: simukulipira mtengo wowonjezera chifukwa cha kukhulupirika pano, choncho masamu a kuyesanso ndi ndalama zaulere. Ngati mtundu wa 3-bit ukanakhala wotsika mtengo pakanakhala kusinthana kwenikweni koyenera kulinganiza — Hy3 imalipiritsa chimodzimodzi ndi chinthu chotsika mtengo kwambiri mu katalogi.

07Zilankhulo zitatu za API, mtundu umodzi

Hy3 imafikika pa zilankhulo zonse zitatu za gateway ndi streaming pa iliyonse. Mtundu womwewo, zolemera zomwezo, malire omwewo — sankhani chimene chikugwirizana ndi SDK imene muli nayo kale.

EndpointChineneroGwiritsani ntchito pamene
/v1/chat/completionsOpenAIMuli ndi SDK ya OpenAI, LangChain kapena LlamaIndex. Lozani base URL kuno, sinthani id ya mtundu.
/v1/messagesAnthropicMuli pa SDK ya Anthropic ndipo mukufuna zigawo za zinthu popanda wosanjikiza wa kumasulira.
/v1/responsesResponsesMukufuna mawonekedwe atsopano a pempho ndi zotuluka za mtundu wodziwika ndi mkhalidwe wa mbali ya seva.

Hy3 imathandizanso kuyitanitsa ntchito ndi kuganiza, kuphatikiza reasoning_effort. Pa ntchito yotulutsa deta njira yothandiza nthawi zambiri ndikutsitsa kuganiza: pa schema yofotokozedwa bwino pa mawu oyera, ma token a kuganiza ndi ma token otuluka amene mumalipira ndi kuchedwa kumene mumadikirira, pa phindu lochepa. M'malo mwake perekani ndalama za khama pa chilowetso chosamveka bwino zoonadi.

Mndandanda wonse wa mapempho ndi mayankho, mindandanda ya ma parameter, tanthauzo la streaming ndi malo oyeserera ogwirizana zili mu zolemba za API.

08Kusankha pakati pa mitundu itatu ya W8A16

Mitundu itatu ili pa mlingo wa kukhulupirika kwapamwamba kwambiri, yonse pa mtengo wa pansi. Chisankho makamaka ndi cha mtundu wa zolowetsa:

MtunduMasomphenyajson_schemaMtengo kulowa / kutulukaMusankheni pa
Hy3 · W8A16AyiInde$0.50 / $2.00Njira zokonzedwa za mawu okha, kulinganiza zida, kubwezeretsa pa context yaitali
MiMo-V2.5-Pro · W8A16AyiInde$0.50 / $2.00Ntchito ya mawu okha kumene mukufuna kukonza kwa Pro kwa mzere wa MiMo
MiMo-V2.5 · W8A16IndeInde$0.50 / $2.00Chilichonse chokhala ndi zithunzi — zikalata zosindikizidwa, zithunzi za skrini, zithunzi

Ngati gawo lililonse la chilowetso chanu ndi chithunzi chisankho chapangidwa kale m'malo mwanu: Hy3 ndi ya mawu okha — tumizani ma PDF osindikizidwa, zithunzi za skrini ndi zithunzi ku MiMo-V2.5. Kumene chilowetso ndi mawu enieni, mbiri ya Hy3 ya kugwiritsa ntchito zida za agentic, kukumbukira kwa context yaitali ndi kupeka kotsika konenedwa imagwirizana mwachindunji ndi ntchito ya njira za deta. Popeza zitatu zonse zimalipiritsa chimodzimodzi, yendetsani schema yanu pa ziwiri za izo pa mbiri zenizeni mazana angapo ndi kuyerekezera milingo ya kuwerengedwa; icho chimathetsa nkhaniyi bwino kuposa tebulo lililonse, kuphatikiza ili.

09Kumene Hy3 imapeza malo ake

  • Kutulutsa magawo kwa kuchuluka kwakukulu. Mapangano, zikalata za boma, malipoti a zochitika, zolembedwa, zolemba za zachipatala, zolemba za makhoti — kulikonse kumene chinthu cha mtundu wodziwika pa chikalata chilichonse ndicho chotulukacho ndipo kuchuluka kumapangitsa kuyesanso kukhala kokwera mtengo.
  • Kuika m'magulu ndi kutumiza pa mlingo waukulu. Zotuluka za enum zotsekedwa, kumene phindu lonse la mtunduwo ndi kusapanga gulu la khumi ndi chisanu ndi chinayi konse.
  • Zolemba za chitetezo ndi zidziwitso za zoopsa. Zolemba za malware, kuwululidwa kwa zolakwika ndi zitsanzo za phishing — zinthu zomwe zimapangitsa mitundu yosefedwa kukana, ndi kumene bowo la mawonekedwe a gulu limawononga cholinga.
  • Ma agent oyitanitsa zida amene sayenera kuima. Kukhulupirika kwa ma argument kwa 8-bit kumatanthauza kuti JSON yoperekedwa ku ntchito zanu imawerengeka; kuyitanitsa kwa zida kosokonezeka ndi agent woimitsidwa, osati yankho loipa pang'ono.
  • Kusanthula kwa zikalata zazitali kupita ku mawonekedwe. Chikalata chonse kulowa, chinthu chimodzi chachifupi kutuluka.
  • Kuyang'anira zinthu ndi kafukufuku wa ndondomeko. Kuika zilembo pa zinthu zimene mtundu wosefedwa umakana kuwerenga ndi vuto lokhalitsa la trust-and-safety; mtundu wosasefedwa wokhala ndi schema yolimba umalithetsa mwachindunji.

Ndipo kumene siiyenera: chilichonse chokhala ndi zithunzi (gwiritsani ntchito MiMo-V2.5), ndi uinjiniya wa mapulogalamu wa mafayilo ambiri wodziyendetsa, kumene matebulo a Tencent okha akuika Hy3 kumbuyo kwa GLM-5.2 pa SWE-bench Verified ndi kumbuyo kwambiri pa DeepSWE.

Onani magwero

Chonena chilichonse cha kamangidwe ndi benchmark chokhudza maziko a Hy3 chimachokera ku zolemba zofalitsidwa za Tencent kapena nkhani zolemba za izo; mfundo za katalogi zimachokera ku /v1/models yathu yeniyeni. Tsimikizani zonse ziwiri.

Ziwerengero za benchmark ndi zonenedwa ndi Tencent, kuphatikiza manambala a opikisana nawo; kutsimikizira kodziyimira pawokha kunali kukudikirabe pa nthawi yolembera. Ziwerengero za mtengo zawerengedwa kuchokera ku mitengo yathu yofalitsidwa ya $0.50 / $2.00. Palibe nambala pa tsambali yomwe ndi benchmark yoyendetsedwa ndi Shannon kapena kuyerekezera.

10Mafunso ofunsidwa kawirikawiri

Kodi API yosasidwa ya mtundu wa Hy3 ndi chiyani?

Ndi endpoint ya Shannon AI Gateway ya id ya mtundu Hy3-W8A16 — maziko a Hy3 a zolemera zotseguka a Tencent otumikiridwa pa gulu lathu la ma GPU pa kulondola kwa zolemera kwa 8-bit, popanda wosanjikiza wa kukana ndi popanda kusefa kwa zotulutsa. Imapereka zenera la context la ma token 262,144, ma token 16,384 otuluka pa pempho lililonse, kuyitanitsa ntchito, kuganiza ndi reasoning_effort, ndi zotuluka zokonzedwa zathunthu kudzera mu zonse json_schema ndi response_format, pa /v1/chat/completions, /v1/messages ndi /v1/responses ndi streaming.

Kodi Hy3 imathandiza JSON schema ndi zotuluka zokonzedwa?

Inde. Hy3-W8A16 imathandiza zonse response_format ya njira ya chinthu cha JSON ndi json_schema ya ma schema olimba a mtundu wodziwika — zotuluka zokonzedwa zathunthu, osati gawo lokha. Sizili choncho pa mtundu uliwonse mu katalogi: Laguna-S-2.1 siithandiza konse, ndipo inkling imathandiza response_format koma osati json_schema. Ngati njira yanu ya deta imadalira pangano la mtundu wodziwika, Hy3 ndi umodzi mwa mitundu imene ingalilemekeze.

Kodi W8A16 amatanthauza chiyani, ndipo ndi wabwino kuposa 4-bit kapena 3-bit?

W8A16 amatanthauza zolemera za 8-bit ndi zochitika za 16-bit — mlingo wa kukhulupirika kwapamwamba kwambiri umene timatumikira. Poyerekeza ndi W4A16-AUTOROUND-REAP (zolemera za 4-bit, kuzungulira kophunziridwa, kudula akatswiri) ndi 3BIT-REAP (zolemera za 3-bit kuphatikiza kudula akatswiri), umasintha zolemera zofalitsidwa pang'ono kwambiri, choncho khalidwe limakhalabe pafupi kwambiri ndi checkpoint yoyambirira. Milingo yolimba ilipo chifukwa maziko akulu kwambiri akanakhala osatheka kutumikira mwachangu.

Kodi Hy3 imathandiza zithunzi kapena masomphenya?

Ayi. Hy3-W8A16 ndi ya mawu okha. Kutumiza zigawo za zithunzi kwa iyo sikungagwire ntchito. Ngati mukufuna masomphenya kuphatikiza zotuluka zokonzedwa pa mlingo womwewo wa 8-bit, gwiritsani ntchito MiMo-V2.5-W8A16, imene imathandiza zolowetsa za zithunzi pamodzi ndi json_schema ndi response_format pa mtengo womwewo wa $0.50 / $2.00.

Kodi API ya Hy3 imalipiritsa zingati?

$0.50 pa miliyoni ya ma token olowa ndi $2.00 pa miliyoni ya ma token otuluka — mtengo wa pansi mu katalogi yathu, wogawana ndi MiMo-V2.5, MiMo-V2.5-Pro, DeepSeek-V4-Flash-0731 ndi MiniMax-M3. Njira ya deta yoyimira yotulutsa magawo pa ma token 1,200 olowa ndi 300 otuluka pa yitanidwe imalipiritsa pafupifupi $0.0012 pa yitanidwe, kapena pafupifupi $1,200 pa mayitanidwe miliyoni.

Kodi Hy3 idzakana pempho?

Palibe wosanjikiza wa kukana ndipo palibe kusefa kwa zotulutsa patsogolo pa mtunduwo. Pa njira yokonzedwa ya deta icho ndi khalidwe la kudalirika mofanana ndi la zinthu: kukana ndi kuphwanya kwa schema, chifukwa ndime yofotokoza chifukwa chake mtunduwo sudzayankha siiwerengeka ngati chinthu chimene chotsimikizira chanu chikuyembekezera. Kuchotsa njira ya kukana kumachotsa gulu lonse la zolephera za kuwerenga pa zolemba za chitetezo, zachipatala, zamalamulo ndi za akuluakulu.

Yendetsani Hy3 pa schema yanu

Kukhulupirika kwa 8-bit, json_schema yolimba, context ya 262K, palibe kukana — pa mtengo wa pansi.

Werengani Zolemba za API Kafukufuku Wonse wa Mitundu

Id ya mtundu Hy3-W8A16 · mawu okha · $0.50 kulowa / $2.00 kutuluka pa ma token 1M


Magwero: Makaadi a mitundu a Tencent Hunyuan · Malo osungira a Hy3 · Hunyuan-Large: An Open-Source MoE Model (arXiv 2411.02265) · Nkhani ya VentureBeat ya kutulutsidwa kwa Hy3. Ma benchmark a maziko ndi onenedwa ndi Tencent ndipo ndi osakhazikika akudikira kutsimikizira kodziyimira pawokha. Shannon AI imayendetsedwa ndi Shannon Lab LLC, New Mexico, USA.

All research links